Design of a multi-layered privacy-preserving architecture for secure medical data exchange in cloud environments

Abstract

Cloud-based e-health systems make medical data widely accessible for treatment, teleconsultation, and analytics; however, the moment clinical records are pushed to an external cloud, they are exposed to a combination of threats: curious insiders, colluding storage providers, and silent data tampering. Most existing attribute-based encryption (ABE) or standalone anomaly detection solutions address these threats in isolation. To address this gap, this work presents a unified, lifecycle-oriented security framework that integrates Server-Aided Revocable Attribute-Based Encryption (SR-ABE) to enforce fine-grained, revocable access control without requiring the re-encryption of historical data. The cryptographic layer is further strengthened through haze optimization-guided key generation, which enhances key-space exploration and randomness during the encryption process. Data integrity is ensured using a SHA-256–based verification mechanism applied at both the storage and access stages. In addition, an intelligent monitoring layer based on a heterogeneous Mixed Graph Neural Network (MGNN) model interacts with patterns among users, devices, and resources to enable continuous event-driven anomaly detection within the system”s operational latency bounds. Under this integrated design, the proposed model attains 99.1% training accuracy and 96.7% testing accuracy, converges with very low errors (training loss 0.0126, testing loss 0.0357), executes faster than prior approaches with an execution time of 10.4 ms at 50 epochs, sustains a high service throughput of 852–901 kbps even when multiple users are active, operates with reduced energy consumption of 0.214 J, maintains low access latency of 6.6 ms, and raises core security indicators to 99.12% data confidentiality and 98.53% data integrity. This shows that combining SR-ABE, HO-based key generation, cryptographic integrity checks, and MGNN-driven surveillance in a single pipeline delivers both stronger privacy guarantees and better runtime performance than existing cloud e-health security schemes.

Introduction

The modern method of patient monitoring using telecare systems is cloud-based and typically uses mobile phones with Bluetooth and cloud infrastructure. Such telecare systems would permit the very early and immediate collection of clinical parameter data using a sensor network for cardiac monitors, which could then be securely uploaded to cloud servers for inspection by remote users. The cryptographic proxy re-encryption protocol for deterring unauthorized access and ensuring compliance with the Health Insurance Portability and Accountability Act will help patients in the dynamic management of access rights to their health information. This system allows efficient revocation for actual deployment in healthcare networks, which also guarantees confidentiality, integrity, and scalability¹. Medical images and Electronic Health Records have seen a wave of extensions in the Peer-to-Peer Virtual Cloud framework, which provides a more extensive application of security for storage processes and retrieval of 3D medical images. The entire protection is guaranteed by the wide adoption of Web-Service Security (WS-Security)-related protocols on encryption and authentication, with key bindings to the user credentials for Structured Query Language-aware encryption of CryptDB. In contrast, a lossless compression and watermarking module minimizes bandwidth consumption without compromising the quality of the images while allowing easy access to clinical data in secure and device-independent ways².

The purpose of establishing a cloud-supported Wireless Body Area Network model is to manage the increasing biomedical multimedia flow, improve confidentiality, and enhance the quality of service. The four-level structure, comprising perception, network, cloud, and application, handles the sensing, transmission, encrypted storage, and access control. Communication is reliable through ZigBee and Transmission Control Protocol/Internet Protocol (TCP/IP), whereas redundancy and energy waste are avoided using content-centric routing and an adaptive QoS mechanism, which results in fast and secure health data transfer between the sensors and cloud services³. One of the solutions developed to overcome the performance and security limitations in healthcare clouds is an advanced multimedia medical-archive system. The model uses the power of Compressed Sensing, Chaotic Multi-Image Encryption, Identity Mutual Recognition Keys and Three-Dimensional Positioning Message-Digest-5 hashing algorithm. Multimodal patient data were aggregated with a two-dimensional (2D) image cube and encrypted with M IE and dual patient and institutional authorization. The 3DPMD5 hash also provides anti-tampering verification, resulting in a secure and scalable ecosystem spread across the edge and central clouds⁴. This is accompanied by threats to privacy and interoperability with centralized electronic health record (EHRs) migration to the digital fold. A decentralized storage, IPFS blockchain-based design was developed.

Encrypted medical files are stored off-chain in the IPFS, whereas on-chain entries only log the pointers and access permissions of these files. Patients own the encryption keys and can provide their clinicians with time-limited access. This architecture guarantees immutability, confidentiality, and transparency in auditing, preventing any single-point failures in managing healthcare data⁵. The availability of IoT wearables increases healthcare data but also enlarges the attack surface. A dual-layer model that blends predictive analytics with strong encryption to secure such systems was proposed. This uses an inception V-3 feature extractor heart-disease detection module in conjunction with an extreme learning machine (ELM) optimized using hydro binary particle firefly optimization (HBPFO). Simultaneously, the Intelligent Encryption and Decryption Framework (IEDF) employs the Advanced Encryption Standard (AES), Data Encryption Standard (DES), Rivest–Shamir–Adleman (RSA), and Modified Blowfish (MBF) algorithms. Keys are randomized in design, ciphers are alternated on a per-block basis, and the design implements a trigonometric key-dependent substitution box, presenting therefore a high degree of precision and resilience to cryptographic attacks⁶.

The security and confidentiality of health data are assured in the CloudLock model for authenticated encryption, such as using ChaCha20-Poly1305 and ECDH-based key agreement, and offers some guarantees in terms of confidentiality and integrity of healthcare data. Minimum guarantees will be enforced, with only authenticated users allowed to decrypt the data with some light overhead. It claims to provide maximum resistance to data loss by splitting sensitive data into small randomly distributed pieces across independent clouds, thereby increasing the difficulty of achieving data loss and improving access control under health security protocols⁷. Such cryptographic mechanisms are structured into attribute-based encryption and secure key exchanges, which include sophisticated key management and authentication, recently adopted in multi-user health systems. Confidentiality guarantees of the data during transmission and storage and efficiency in user verification have thus been included in the model. This proves its application as a flexible framework to prevent unauthorized access and accidental alteration, making it a flexible solution for secure communication in healthcare cloud networks⁸.

The integration of CPEBLD with blockchain auditing and convergent encryption allows for the non-redundant storage of EHRs, enhancing confidentiality in the United States. This approach removes any redundancy of data blocks used on different patients and ensures that the data remain confidential using smart contracts and personal verification. Integrity evidence in the blockchain prevents interference by third parties in auditors, and the scheme is spatially efficient, calculability-tractable, and resistant to forgery attacks to manage the records of doctors in a secure and scalable way⁹. Since then, the gradual evolution of Healthcare 4.0 to an adjacent hybrid cryptographic architecture that combines the advantages of ECC and blockchain has been guided. It is based on the architecture of four overlapping layers: edge, fog, cloud, and blockchain, such that biomedical multimedia molecules (X-ray, CT, and MRI data) are delivered and stored securely in the cloud. This secret key exchange is carried out using ECDH and authentication using ECDSA. The blockchain stores encrypted metadata and audit logs to be tamper-proof, quantum-safe, low-computation-time, and high-image-fidelity to guarantee integrity and traceability¹⁰.

Emerging technologies such as cloud and IoT are preparing to build a patient space that is constantly monitored, where large data storage and transfers are made, and intelligent decisions are made in health clinics. In addition, the numerous infrastructures bring problems in terms of the interoperability of different cloud environments, the absence of integration between encryption and anonymization and a mechanism with access control, and the augmented threats of unauthorized access, data spoiling, and identity innuendo. All these difficulties are largely explained by the fact that hospitals are likely to archive blocks of medical information with diagnostic images and electronic health records using interconnected clouds. Hence, it is necessary to develop a complete adaptive and computer-efficient security model that will ensure confidentiality and trust in a patient but enable very precise measurements and easy flow of information that will support evolving regulatory demands in relation to privacy in healthcare. This serves as a prima facie motivation for the next-generation architecture on the cloud with privacy protection, maintaining security without harming usability in actual working medical settings.

To date, existing frameworks in the area of cloud healthcare have made significant advancements in data protection schemes, such as homomorphic encryption, federated learning, blockchain, and access control models. However, the solutions are mostly domain-specific and either approach data confidentiality alone or model performance in isolation. The architecture of the current generation suffers from failure to adapt in real time over distributed environments on one hand or settling for a trade-off between privacy preservation, computational cost, and interoperability on the other hand. In addition, multi-source healthcare data from wearable IoT sensors to imaging and genomic repositories have no cohesive structure for anonymizing and encrypting messages; therefore, they have optimal leakage during analytics or transmission. Therefore, a critical gap exists in end-to-end, intelligence-driven security frameworks incorporating dynamic encryption, optimized anonymization, and federated decision-making at high throughput and diagnostic viability in heterogeneous healthcare cloud ecosystems. Based on the aforementioned research gap, this work makes the following contributions:

• A comprehensive end-to-end security framework was designed for cloud-based healthcare systems, ensuring that every phase, from data creation to monitoring, is protected through SR-ABE encryption, integrity validation, and intelligent surveillance.

• The Server-Aided Revocable Attribute-Based Encryption (SR-ABE) system can offer partial decryption to the cloud, prevent the exposure of the plaintext, and release effective key revocation that does not require re-encryption of the entire dataset.

• Haze optimization integration in the key-generation process maximizes key entropy, producing highly random and resilient cryptographic keys that resist brute-force and correlation-based attacks.

• An integrity-checking layer based on SHA-256 ensures that every encrypted record is authenticated by a fingerprint that is tamper-evident to prevent any attempts at corruption or alteration of data, which can be easily detected during retrieval.

• The Mixed Graph Neural Network (MGNN) module represents user, device, and file relationships on the healthcare cloud to identify abnormal access behaviors, insider threats, and collusion patterns, which cannot be detected by conventional intrusion detection systems.

Although encryption, integrity verification, and intrusion detection have been individually explored in cloud-based e-health security, this work introduces a fundamentally different end-to-end, lifecycle-driven security architecture in which these mechanisms are tightly coupled rather than independently deployed. The proposed framework uniquely integrates Server-Aided Revocable Attribute-Based Encryption (SR-ABE) with an optimization-guided high-entropy key generation process, enabling scalable fine-grained access control with efficient user revocation while simultaneously strengthening resistance to brute-force and key-compromise attacks, an aspect not addressed in conventional ABE-based schemes. In addition, unlike existing approaches that apply integrity checking only at the storage level, this work incorporates dual-phase SHA-256–based integrity validation at both the data upload and access stages, ensuring the detection of silent tampering during data transit and retrieval. Furthermore, the security architecture extends beyond cryptographic protection by embedding a heterogeneous Mixed Graph Neural Network (MGNN) that explicitly models relational interactions among users, devices, and cloud resources, allowing real-time detection of complex multi-entity attack patterns that cannot be captured by traditional rule-based or flat-feature machine learning detectors. By unifying cryptographic enforcement, integrity assurance, and intelligent behavioral surveillance within a single coordinated pipeline, the proposed approach addresses the critical challenges of scalability, revocation efficiency, and adaptive threat awareness that remain unresolved in existing cloud e-health security solutions.

The remainder of this work is structured as follows. Section “Related works” summarizes the current cloud healthcare structures. Section “Overview of the proposed framework” defines the hybrid architecture with the highest level of privacy and explains the workflow of the system, algorithmic elements, and security integrations. Section “Decryption phase” covers the experimental setup, implementations, and evaluation metrics, which are the performance, scalability, and efficiency of data protection. They conclude with a summary in Section “Revocation phase”, which addresses future research directions to achieve adaptability, scale down the complexity of the computation, and deliver safe real-time medical information exchange between heterogeneous clouds.

Related works

Digital healthcare has triggered a transformation in the manner in which clinical data are created, stored, and shared. EHRs are cloud systems that facilitate disruption in the healthcare industry by enabling a safe environment for handling and storing massive amounts of information. Cloud computing is scalable; therefore, it can combine diagnostic imaging, genomics, and patient history in the same system that can be accessed by authorized users wherever they are. Some cybersecurity risks related to sensitive health information include ransomware, unauthorized access, and even data corruption during their transfer to off-location servers. To overcome such threats, researchers have implemented powerful encryption tools, audit trails, and adherence to HIPAA and GDPR, among other laws. In addition, a new concept of cloud biosecurity has been presented, whereby bioinformatics security and cybersecurity converge to deliver superior collaboration and results to patients, as well as guarantee the confidentiality, integrity, and availability of data¹¹.

The suggested privacy-preserving analytical architecture enables the manipulation of sensitive healthcare data while preserving the utility of these data. The model builds on the concept of zero-knowledge proofs, namely Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge (zk-SNARKs) in the multi-tenant cloud setting implemented on blockchain. Anonymized raw health data are then converted to privacy-preserving parameters, and zk-SNARKs are used to verify computations without revealing the underlying information. They can perform two possible operations: the first is performed off-chain, and the second is a cryptographic proof stored on a personal blockchain ledger, and the two are synchronized through a smart contract, which is used to secure the sharing and access of data. The experimental data tests with synthetic telemedicine showed high levels of anonymity, computer efficiency, and regulatory adherence, which proved the framework to be a scalable and privacy-neutral healthcare analytics solution¹².

Ciphertext-Policy Attribute-Based Honey Encryption (CP-ABHE) was introduced to address large-scale healthcare data security challenges. A dual-layer encryption infrastructure through honey encryption is characterized by Ciphertext-Policy Attribute-Based Encryption (CP-ABE), which permits fine-grained attribute-level access control and generates plausible decoy data for incorrect passwords, preventing brute-force and side attacks. In Hadoop/HDFS environments, data are encrypted during write operations, whereas credential verification occurs at read times for secure MapReduce processing. Assessment results obtained using Google Cloud infrastructure demonstrated that the new system reduced encoding and decoding time, increased the amount of data processed, and performed better than standard algorithms such as DES, RSA, Blowfish, AES, and AES-OTP, while being robust against inside and man-in-the-middle attacks¹³.

Another work presented integrated IoT-cloud health management systems and Elapid Encryption (EE), which was proposed for transmitting data securely over the Internet, and Generalized Fuzzy Intelligence–Ant Lion Optimization (GFI-ALO) as a classifier for disease prediction and severity assessment. IoT sensors continuously monitor physiological indicators, such as heartbeats, blood pressure, and cholesterol levels, and securely send the data through EE to the cloud. Disease classification of patients was performed excellently using fuzzy logic and optimization-based learning using the GFI-ALO model. The experimental validation showed that this can provide more accuracy, more precision, better recall, and secrecy in the results with less computation time and error rate, thus, it is flexible for real-time healthcare analytics, even in remote or rural sites¹⁴.

The aforementioned Scalable Subtree L-Anonymization (SS-LA)- based cloud-centric privacy model integrated with the Whale Optimization Algorithm (WOA) is targeted to ensure privacy over large EHR datasets. After data normalization and feature extraction based on ICA, SS-LA generalized the quasi-identifiers by performing hierarchical operations such as suppression, bucketization, and randomization, while the WOA mechanism optimized the utility of the data with consideration of privacy protection. A series of experimental comparisons with more conventional anonymization-based techniques proved its superiority in terms of execution time, accuracy, and F1-score, thus suggesting model scalability and capability in the development of a secure cloud-based EHR management system¹⁵. The Secure Electronic Health Record-Blockchain-Cryptographic Hash Generator (SEHR-BC-CHG) was proposed to strengthen the integrity and confidentiality of cloud-based EHRs. It was combined with a Cryptographic Hash Generator (CHG) and blockchain consensus based on proof-of-monitoring (PoM) and Discrete Shearlet Transform (DST) encryption of sensitive fields.

In addition, a Hybrid Chaotic Atom Search-Tree-Seed-Levy Flight (HCASOA-TSA-LF) optimizer was used for the validation of access requests and anomaly detection. SEHR-BC-CHG achieved reduced encryption and decryption speeds, reduced computational load, and was superior in terms of reliability, throughput, and confidentiality compared to other SEHRs¹⁶. This is a cloud-enhanced security solution and is commonly referred to as the Dual Kernel Support Vector machine Crossover Wild Horse (DKS-CWH). It was suggested that the storage and categorization of digital images in the medical profession should be enhanced. It uses a dual-kernel SVM to extract features and tune its parameters using the Crossover-based Wild Horse Optimizer (CWH) to avoid local optima. The system can be used to process medical images with enhanced results in terms of accuracy, precision, recall, and F1-scores, and reduce the execution time and misunderstanding in the medical imaging case at credible and real-time cloud-based image analysis when using the Medical MNIST dataset¹⁷.

The other innovation is a Blockchain-based Chaotic Tent Map Encryption Scheme (BCTMES) that prevents the security of medical images in the cloud conduit and stored images. This method employs the immutability aspect of the blockchain with chaotic tent maps and prime circulant matrices to offer high permutation and diffusion properties. The experiments showed a high level of entropy and low level of correlation, equal histograms, and high resistance to brute-force and differential attacks, which provide a security structure of confidentiality, integrity, and availability of data in cloud medical imaging workflows¹⁸. S2Cloud was created to implement the safe handling of massive amounts of health data streams created by the Internet of Things. It connects wearable sensors and smartphones to the cloud in real time via a pipeline to assist in the visualization and tracking of long-term physiological data. The schema incorporates user verification with AWS Cognito, where authenticated clinicians are the only users who can access patient information. Owing to its modularity, physicians will be able to communicate and interact with patients in real time, analyze live signals, and handle patient records, thereby making it a scalable telehealth solution of the future¹⁹.

The Secure Healthcare Access Control System (SHACS) permits access to public cloud access control systems based on conventional Role-Based Access Control (RBAC) with a twist by incorporating Attribute-Based Access Control (ABAC), multi-factor authentication, and anomaly detection. SHACS allows the accurate authorization of the context and quick identification of unauthorized actions through the enforcement of adopted rules and observation of the behavior patterns of users. The outcomes of the experimental studies revealed the following: the access latency was shorter, scalability was better, and protection against unauthorized access increased²⁰. Homomorphic encryption has been used to protect data introduced into cloud EHR systems, such as information. A real encryption system using homomorphic encryption was constructed that calculates encrypted data without affecting privacy, particularly during processing and transfer. It uses fully and partially homomorphic encryption schemes, cryptographic hashing, and signature validation to ensure privacy, integrity, and interoperability among medical institutions²¹.

The HCPMP hybrid cloud performance management model encourages effective patient monitoring by connecting IoT wearable sensors and cloud analytics. In this framework, data encryption was performed using Elliptic Curve Cryptography (ECC), and various machine learning classifiers, namely Random Forest, SVM, Decision Tree, Logistic Regression, XGBoost, and Gradient Boosting, were applied to facilitate the real-time detection of health anomalies. With evaluations having been carried out on real datasets, the results indicated high classification accuracy with low latency and even some level of resilience against hacking, thus marking HCPMP as a good architecture for remote healthcare monitoring²². Privacy-Preserving Federated Learning with Homomorphic Encryption (PPFLHE) addresses the issue of collaborating on medical AI without sharing raw patient data. Each institution trained a local EfficientNet-B0 model under the CKKS homomorphic encryption scheme, and a central server performed encrypted gradient aggregation using federated averaging (FedAvg)²³.

For anomaly detection in cloud-based medical Internet of Things (IoT) networks, an Explainable AI (ExAI)-driven model based on a Radial Boltzmann Gaussian Temporal Fuzzy Network (RBGTFN) and Remora Colony Swarm Optimization (RCSO) algorithm was proposed. Networks such as DARPA, CAIDAS, and DEFCON enable experimentation, demonstrating improvements in the current model over classical machine learning models in terms of detection accuracy, latency, and quality of service²⁴. The following work introduces a safe and effective medical image encryption paradigm suitable for healthcare IoT applications that utilizes a recently developed three-dimensional hyperchaotic map (3DLCSA) hybridized with the integer wavelet transform (IWT) and advanced statistical operations utilizing DNA. The algorithm takes the approximation bits of the image using IWT to minimize the data space and calculation cost, and then multi-stage encryption using diffusion, bit-level permutation, and randomized DNA coding by DNA cubes of specially designed DNA²⁵.

This high-performance IoT-cloud framework, including an adopted custom Spec Trasafe Algorithm, Facile Hash Algorithm, and Dynamic k-anonymity (DKA Algorithm), seeks to protect healthcare information during the resting and transportation procedures. Several performance comparisons against this high-quality data metric prepared on the basis of Kaggle were made to compare it with the existing methods SMPC (Secure Multi-Party Computation), MSHC (Multi Server Homomorphic Encryption), AHC-ASS (Attribute-Hiding Cryptography with Attribute-Based Secret Sharing), and MP-FHE (Multiparty Fully Homomorphic Encryption)²⁶. Secure telemedicine platforms for this cross-institution EHR data sharing were then built into a cloud-based architecture of Software, Infrastructure, and Platform-as-a-Service layers. It harmonizes electronically created health records as HL7 CDA and ICD coding for interoperability with mixed centralized-decentralized storage and high-security APIs. This framework managed 99.5% data transfer success because of a mere 0.2% loss vulnerability rate, which stands as a testimony to its effectiveness in large-scale compatible medical data exchange²⁷.

Both Attribute-Based Encryption (ABE) and Attribute-Based Access Control (ABAC), along with searchable encryption and Revocable Searchable Attribute-Based Encryption (RSABE), were accommodated in the next generation of EHR frameworks using a semantic knowledge graph architecture²⁸. In this way, the patient records were modeled as encrypted graph nodes, which improved the query precision and adaptiveness while reducing the latency speed by approximately 40% to ensure scalability for large healthcare datasets. A secure telemedicine system combining IoT sensors, home gateways, and a cloud application endpoint employs Transport Layer Security (TLS) 1.3, Elliptic Curve Diffie-Hellman Ephemeral (ECDHE), Elliptic Curve Digital Signature Algorithm (ECDSA), Advanced Encryption Standard in Galois/Counter Mode (AES-GCM), and RSA-4096 for end-to-end confidentiality and authentication²⁹. The system achieved transmission times of 186–188 ms with high reliability under loads to ensure real-time patient monitoring over the Internet. The final configuration is a hybrid Internet of Medical Things (IoMT)-blockchain-edge computing framework for performing decentralized data management of health information³⁰. Wearable sensors collect physiological data that are pre-processed at the barrier of latency. Important metadata were anchored to the Hyperledger Fabric blockchain for traceability and auditability. This entire setup maximizes the throughput, synchronization, and network stability of an intelligent healthcare system infrastructure that is scalable and privacy-aware. Table 1 presents a comparative analysis of the existing techniques applied to cloud-based healthcare systems.

Table 1.

Comparative review of existing cloud-based healthcare data security models.

Authors	Methodology	Advantages	Limitations
Thilakanathan et al.1	Proposed a Cloud-based telecare model using smartphone and proxy re-encryption for secure health data sharing	Ensures secure and confidential data exchange between patients and doctors	Lacks detailed quantitative performance evaluation metrics
Castiglione et al.2	Lossless 3D predictive compression with LSB watermarking delivered via a dynamic peer-to-peer “Virtual Cloud.”	Secure, adaptive, device-agnostic access with strong crypto and improved BPP at N = 4	Many network-level and ML metrics not evaluated or reported
Hassan et al.3	Cloud-assisted WBAN using NDN/CCN and adaptive streaming	Enables low-latency, QoS-aware medical data sharing	Energy efficiency and security aspects not deeply analyzed
Zhang et al.4	Cloud-based multimedia health archive using CS, MIE, IMRKs, and 3DPMD5	Achieves efficient compression, strong encryption, and cloud anti-tampering	Does not compress 3D models; limited data-type compatibility
Jeyanthi et al.5	Patient-centric blockchain with off-chain IPFS, AES for data and RSA envelopes for key sharing	Greatly reduces on-chain size and preserves privacy with patient-controlled access	No empirical latency/throughput or energy results; retrieval gains shown only comparatively
Pichandi et al.6	HybBPF-ELM for prediction with IEDF (AES/DES/RSA/Modified-Blowfish + ASC) for cloud security	High diagnostic accuracy with faster, layered cryptography on cloud data	Computational load, class imbalance, and generalisability across more datasets need work
Chitra et al.8	Introduced the Improved Diffie–Hellman Key Exchange Algorithm (IDHKE) integrated with attribute-based encryption for secure key sharing and controlled access in healthcare cloud systems	Provides stronger confidentiality and integrity with authenticated key exchange and efficient resource use	Does not address energy metrics or real-time performance factors like latency and throughput
Vivekrabinson et al.9	Proposed a Cross-Patient Encrypted Block-Level Deduplication (CPEBLD) model using blockchain auditing and convergent encryption to securely store and verify EHRs without third-party auditors	Achieves secure, decentralized integrity verification, reduced redundancy, and minimal computation on clients	Does not evaluate energy metrics or predictive accuracy parameters such as precision or F1 score
Junnarkar et al.10	Developed a Healthcare 4.0 system using integrated lightweight ECC and private blockchain for secure biomedical multimedia transmission	Ensures faster encryption/decryption, higher image quality, and strong data integrity without third-party dependence	Energy-related and network-level metrics were not analyzed
Sachdeva et al.11	Integrated cloud computing with biomedical data workflows for secure storage and analysis across EHRs, genomics, proteomics, and clinical systems	Enables scalable data handling, remote access, and collaborative healthcare delivery while ensuring compliance	Lacks experimental evaluation and quantitative performance metrics for proposed approaches
Babu et al.12	Integrated zk-SNARKs and blockchain in a multi-tenant cloud to secure healthcare analytics	Provides strong privacy, transparency, and efficient computation for sensitive health data	Relies on synthetic data and faces performance overhead for large-scale deployments
Kapil et al.13	Hybrid CP-ABE + honey-encryption embedded in HDFS/MapReduce	Faster crypto throughput with decoy-based defense and fine-grained access	Usability, latency, and compliance considerations noted by authors
Verma et al.14	IoT-cloud healthcare system with Elapid Encryption and GFI-ALO classifier	High accuracy, security, and classification performance	Energy use and scalability not analyzed
Shanthi et al.15	SS-LA with WOA after ICA-based feature extraction and normalization for scalable EHR anonymization in cloud	Cuts memory use and runtime while improving accuracy and F1 in privacy-preserving analytics	Reports limited utility metrics (e.g., no throughput/recall/specificity), so broader trade-offs remain unquantified
Basha et al.16	Blockchain with PoM + CHG, DST encryption, and HCASOA-TSA-LF request optimization for cloud EHR	Cuts overhead and boosts confidentiality and reliability in EHR sharing	Reports many gains as percentages without broad absolute utility metrics
Kokila et al.17	DKS-CWH combines dual kernel SVM with a crossover-based wild horse optimizer for secure medical image classification in the cloud	Achieves high accuracy with lower error and faster execution	Relies on strong network connectivity and faces potential data privacy challenges
Shahid et al.18	Combines blockchain verification with chaotic tent map encryption for secure cloud-based medical image sharing	Offers strong security with high key sensitivity and attack resistance	Computational complexity may impact real-time performance
Stauffer et al.19	Developed a cloud-IoT platform (s2Cloud) for secure, real-time health data management and visualization	Enables continuous data streaming with patient interaction and secure access	Quantitative performance metrics and attack-model testing are not provided
Sangeetha et al.20	Developed SHACS combining RBAC, attribute policies, and anomaly detection for secure EHR access	Reduces authentication time and enhances scalability and resilience	High computational cost limits large-scale deployment
Annapurna et al.21	Designed a cloud-based EHR framework using homomorphic encryption for secure storage, processing, and sharing	Ensures data privacy and supports collaborative analytics	High computational overhead and complex integration
Ebadinezhad et al.22	Developed a secure, real-time IoT healthcare monitoring system (HCPMP) integrating ML algorithms with ECC-based encryption	Achieves perfect classification accuracy and strong security under attacks	Limited evaluation against complex, combined attack scenarios
Adnan et al.23	Developed PPFLHE framework combining federated learning and CKKS homomorphic encryption	Enables secure model training without data sharing	Higher computational overhead and latency
Samriya et al.24	Proposed an ExAI-based anomaly detection model (RBGTFN-RCSO) for securing healthcare IoT networks	Provides high accuracy and transparency in anomaly detection	Increased computational complexity during optimization
Lai et al.25	Proposed a 3D hyperchaotic map with IWT and DNA cubes for secure medical image encryption	Achieves high encryption strength, randomness, and resistance to attacks	Limited to 2D image encryption; future work needed for 3D data
Ammerha Naz et al.27	Developed a cloud-based framework integrating HL7 CDA and hybrid data exchange for global EHR interoperability	Achieves high data transfer success and strong security compliance	Limited performance metrics beyond latency and transfer success
Walid et al.28	Proposed a graph-based EHR system using RSABE, ABAC, searchable encryption, and SPARQL queries for fine-grained access and scalable data handling	Enables field-level security, efficient querying, and supports dynamic attribute revocation	Some performance metrics (like accuracy or energy use) were not evaluated
Sornlert et al.29	Developed a secure, cloud-based telemedicine system integrating IoT devices with multi-layered encryption	Ensures strong data security with minimal latency for real-time monitoring	Some performance metrics (like accuracy or energy) were not evaluated
Abbas et al.30	Integration of IoMT, edge computing, and Hyperledger Fabric for secure, real-time healthcare data validation and storage	Enhances data privacy, reduces latency, and enables trustworthy data sharing	Commit phase introduces latency; large-scale deployment faces integration and cost challenges

Smart medical gadgets and digitized health records produce massive amounts of data that are streamed to the cloud at remarkable speeds. Therefore, in this situation, the possible data being intercepted and the threat from internal employees gain notoriety, now compounded by a huge population sharing a cloud with high computational scalability. Almost all the erstwhile encryption and access control schemes have poorly accommodated situations in multi-cloud and hybrid-edge environments, which extend the responsibility of the data to be owned in a decentralized manner and shift trust regions dynamically. In the absence of mixing such strong security measures, the confidentiality, integrity, and compliance of patient data with protection acts such as HIPAA and GDPR would otherwise be jeopardized. When such times arise, a thick-light-adaptive security framework is most desirable for protecting medical data for their entire lifetime, from collection to transmission to cloud storage and computation. This work examines a hybrid paradigm that incorporates privacy-sensitive concepts and optimizes encryption and smart access controls in a scalable performance-conscious cloud architecture, which will permit the expeditious and consenting exchange of health information prepared for analysis without affecting computational performance and correctness.

Overview of the proposed framework

This framework is an integrated and dynamic architecture that suggests dealing with security problems in medical cloud computing that control health data. It ensures the continuity of the right since the time data are generated to ensure that there is access to the health information of patients at any point. The original workflow can consider the personal health record (PHR) information generation of medical devices, clinical systems, or healthcare staff. Whenever they are generated, they are automatically linked to encryption by the SR-ABE description (Server-Aided Revocable Attribute-Based Encryption). The SR-ABE implements attribute-based access control, where healthcare employees with established credentials and permissions are allowed to decrypt and access files encrypted by this scheme. In addition, the structure uses Haze Optimization (HO) during the key generation stage to produce hard encryption keys. The HO algorithm changes to search a large solution space when generating keys, generating high-entropy cryptographic keys in the process, hence lowering the chances of breaking such keys or exposing them to brute-force attacks. The data are then encrypted, the key is secured, and the data cross through the layer of verification of the integrity, that is, the sha-256 security layer, which generates a distinct cryptographic signature for every dataset. This enables the instant identification of any variation, corruption, or distortion of information during retrieval and storage. The validation of data integrity opens the way to transferring data to cloud storage, which is always encrypted and under surveillance. It also uses a detection system for threats based on an architecture that adheres to a Mixed Graph Neural Network (MGNN) to counter ever-changing cyber threats.

This smart monitoring module observes communication patterns, user attitudes, and access behaviors across different networks. In this manner, it learns relational patterns among entities (users, devices, and data files) to identify anomalies or malicious intent, insider misuse, and unauthorized collaboration attempts. Upon the occurrence of an access request, the framework is set to assess it against a specific decision layer. Granted access requests are then decrypted and fetched for the response, and questionable or unauthorized access requests are rejected in due time. This ensures that each layer contributes independently to the CIA triad, as illustrated in Fig. 1 below:

• Further sustained with SR-ABE encryption and attribute-based access control,

• Enabled real-time monitoring within the cloud infrastructure using the MGNN. Thus, it assures trusted access to data for legitimate users.

Fig. 1.

Workflow of the proposed cloud-based e-health security framework.

Together, these platforms construct a multimodal barrier protecting cloud-based healthcare from internal and external attacks while ensuring trust, privacy, and safety, which are essential in today”s digital health environments.

System architecture and data flow

The proposed architecture provides a secure and intelligent workflow for managing electronic health information in a cloud computing environment. The cooperative interaction among four key entities, namely, the Data Owner (DO), Key Authority (KA), Cloud Server (CS), and Data User (DU), is defined as follows. These components ensure that patient data are protected, available, and audit-ready at every phase of the life cycle. The data flow begins with the patient or health source, flows through encryption and key management, and ends with authorized retrieval (Fig. 2). The Data Owner initiates the entire process and collects and prepares medical IDs from sensors, healthcare systems, or hospital networks. Before any transmission, the data are encrypted using the Server-Aided Revocable Attribute-Based Encryption method. This method effectively transforms raw medical information from readable plaintext into gibberish (ciphertext) figures while embedding an access policy with the ciphertext, stating which user groups can decrypt the data. Decryption relies on secret keys and public parameters generated by the Key Authority, a highly trusted entity responsible for attribute assignment, key generation, and revocation.

Fig. 2.

System architecture and data flow of the proposed SR-ABE-based cloud security framework showing interaction among the Data Owner, Key Authority, Cloud Server, and Data User.

The role of the cloud server as a semi-trusted entity is to provide storage and limited computational support without having access to plaintext data. As encrypted health records reach the cloud, the encrypted data are indexed and stored for later access. When a Data User (e.g., certified clinical professional, expert) requests data, the Cloud Server validates the request and supports partial decryption, a process that safeguards against the disclosure of data to unauthorized individuals while relieving the computational burden from the user side. Only individuals whose attributes meet the policy instilled during encryption will successfully create a symmetric copy of the plaintext file.

The sequence of the work is given below:

1. Data Generation: Patient information is collected through electronic medical devices or health record systems.

2. Encryption: The Data Owner encrypts the dataset using the SR-ABE encryption function:

   1

   where is the public key and denotes the defined access policy.

3. Cloud Storage: The ciphertext is uploaded to the Cloud Server for scalable and secure storage.

4. Access Request: A Data User possessing an attribute set requests data access.

5. Key Verification and Decryption: If (the user”s attributes satisfy the access structure), the ciphertext is partially decrypted by the Cloud Server and fully decrypted by the user:

   2

   where is the user”s private key derived from their authorized attributes.

Any inter-entity communication is performed over an absolutely secure channel with SSL/TLS protocol implementation to safeguard the message against interception and replay attacks. The trust model in this scenario considers the Key Authority as fully trustful, the Cloud Server as honest-but-curious, and both Data Owners and Data Users as legitimate but could be compromised. Under these various trust assumptions, even if one of the entities is compromised, the patient”s secret remains protected because of the cryptographic separation of roles. Therefore, confidentiality, integrity, and availability are achieved in an ecosystem through SR-ABE encryption and regulated key distribution; available ciphertexts and verified keys are decrypted to ensure integrity; and availability is facilitated through its distributed cloud architecture and authorized real-time access. These are considered robust mechanisms for developing a secure and resilient framework for healthcare data management in the cloud.

Server-aided revocable attribute-based encryption (SR-ABE) module

The Server-Aided Revocable Attribute-Based Encryption (SR-ABE) framework forms the basis for the cryptographic techniques in the proposed security model. It features fine-grained access control policies that combine user attributes and encryption policies and allows efficient key revocation without necessitating full re-encryption of stored data. The module consists of five main phases, as shown in Fig. 3: Setup, Key Generation, Encryption, Decryption, and Revocation.

Fig. 3.

Server-aided revocable attribute-based encryption (SR-ABE) process integrating Haze Optimization for secure key generation and revocable access control.

1. Setup Phase

In this initial stage, the Key Authority (KA) sets up the cryptographic environment and creates the public and master parameters for the system. The KA selects two massive prime numbers pand q and calculates the bilinear group of order with generator . A bilinear map is defined, where and are cyclic groups of the same order . The authority then chooses two random elements and defines:

3

4

The public key is broadcast to all participants, while the master key remains secret with the Key Authority.

2. Key Generation Phase

When a Data User (DU) registers in the system, the KA issues a private key based on the user”s assigned attribute set . The KA chooses random values for each attribute and generates the private key components as:

5

Here, is a cryptographic hash function that maps attribute strings into group elements. This attribute-bound key ensures that users can only decrypt data whose access policies align with their assigned attributes.

In this module, Haze Optimization (HO) is used to generate high-entropy key seeds to prevent correlation-based attacks. The entropy score of a generated key is computed as:

6

where denotes the bit probability distribution of . The optimization process maximizes to yield the most unpredictable keys.

3. Encryption Phase

The Data Owner (DO) defines an access structure (typically a Boolean access tree) specifying which attribute combinations grant decryption privileges. The data message is then encrypted as follows:

7

Here, is a random session secret for each encryption instance. The resulting ciphertext is transmitted to the Cloud Server (CS) for storage. The CS, considered a semi-trusted entity, can perform delegated computations but cannot decrypt the ciphertext because it lacks the master secret .

4. Decryption Phase

When an authorized user requests access, the CS assists in partial decryption to reduce the computational overhead on the user”s side. The server computes an intermediate component:

8

and forwards it to the user. The final decryption is performed by the Data User using their private key components as:

9

Decryption succeeds only if the attribute set of the user satisfies the policy defined by the Data Owner. This fine-grained decryption ensures that no unauthorized individual, even with partial access to ciphertext or cloud data, can reconstruct the plaintext.

5. Revocation Phase

User revocation is handled through a server-aided update process, which prevents compromised or outdated keys from being used without re-encrypting all stored data. The KA generates an update key for time period and transmits it to the Cloud Server. The CS then modifies stored ciphertexts as:

10

where is a random scalar used for key refreshing. Active users receive their updated decryption keys , allowing continuous access, while revoked users without lose decryption capability. This time-bound revocation approach ensures forward and backward secrecy without heavy computational cost.

The data Owner allows the key authority to encrypt, upload, and manage an information key. He must also ensure computation in the cloud server without touching the actual plaintext and recover data whenever the Data User satisfies the access policies. These methods maintain confidentiality through encryption, integrity through key validation and controlled updates, and availability through continuous secure access to real users. This module, as shown in Fig. 3, integrates the processes of encryption and decryption along with optimizing the keys into a single workflow, thus ensuring that their communication and storage are completely secure in a cloud-based healthcare system.

Haze optimization (HO) for key generation

In this framework, Haze Optimization (HO) is employed to fortify the cryptographic strength of the key generation process used in the SR-ABE module. Conventional key generation methods, which are usually based on pseudo-random number generators, may sometimes yield low-entropy keys that are vulnerable to brute-force or correlation attacks. By adopting the HO algorithm, an adaptive population-based search was introduced, maximizing the key entropy so that every generated key was maximally random and computationally unpredictable. The main objective of the HO-based key generation process is to lower the predictability and maximize the entropy such that each cryptographic key has maximum randomness. High-entropy keys resist statistical and differential attacks, as it becomes negligible to reproduce or guess the key sequence. Hence, we can define the optimization problem as follows:

11

where denotes the entropy of the candidate key , represents the total number of bits in the key, and signifies the probability of occurrence of each bit configuration.

The optimization process continues until the entropy score approaches its theoretical maximum , ensuring uniform bit distribution.

The workflow of HO-based key generation comprises the following stages:

1. Population Initialization:

A random population of candidate keys is generated, where each key is represented as a bit string of length . These candidates form the initial search space.

2. Fitness Evaluation:

Each candidate key is assessed using the entropy-based fitness function . Higher entropy indicates stronger randomness, and therefore, a higher fitness score.

12

where the normalization factor ensures that fitness values remain between 0 and 1.

3. Haze Update Mechanism:

The Haze algorithm mimics the natural phenomenon of haze dispersion, where particle concentration diffuses gradually toward equilibrium. The update rule for each key is formulated as:

13

Here, represents the best-performing key in iteration , is the convergence factor controlling attraction toward the best solution, and introduces controlled randomness to maintain diversity and prevent premature convergence.

4. Selection of Optimal Key:

After several iterations, the algorithm identifies the candidate key with the maximum entropy and minimum redundancy. The final selected key is then used for encryption and decryption in the SR-ABE framework:

14

This optimal key provides a highly secure seed for the cryptographic process, making the system resilient against brute-force, side-channel, and correlation attacks.

Algorithm 1.

Haze optimization (HO) for secure key generation.

SHA-256-based integrity verification

SHA-256 based integrity verification module is integrated into the proposed framework to maintain patient information authenticity against unauthorized manipulation during data transmission or storage. This component increases the strength of the cryptographic assurance level, which verifies the originality of the stored data before and after conducting access operations. The SHA-256 algorithm always produces a unique and irreversible hash value for any file so that this hash will act like a digital fingerprint of the file itself, and it will change even if one single bit of the data is modified.

The process starts when the Data Owner or healthcare system generates the dataset D. This is how it works before encryption and before uploading: the SHA-256 function gets using the formula:

15

This hash value is then securely stored, either as a tag with the metadata of the encrypted file or logged immediately in a tamper-resistant ledger, such as a blockchain-based audit trail. The encrypted data will be uploaded to the cloud together with its hash value.

During the retrieval phase, when a Data User or authorized entity requests access, the framework recomputes the hash value for the retrieved dataset :

16

After which a direct comparison will be made by the system on stored hash against the recomputed hash. If , then the data integrity is considered having confirmed whether the data has not been tampered, damaged, or wrongfully modified when stored or during transit.But if , it will immediately trigger a security alarm or access denied with a mention of fake data.

This mechanism provides one-way protection, meaning that while it is computationally easy to verify a hash, reconstructing the original data from the hash value is nearly impossible. This property demonstrates that integrity verification imposes a low overhead while providing a strong defense against forgery, replay attacks, and unauthorized edits. The framework thereby facilitates the preservation of the original, traceable, and invariant nature of healthcare records by embedding this verification into the data lifecycle.

Algorithm 2.

SHA-256-based data integrity verification.

Heterogeneous network embedding using mixed graph neural network (MGNN)

A hybrid graph neural network (MGNN)-based Heterogeneous Network Embedding Module is being built to fill the intelligent threat detection layer in the proposed framework, which will eventually cover all the aspects necessary for the identification of anomalies in terms of interactions and unauthorized access of data and insider misuse on-the-fet in real time. The operation of the model largely depends on a heterogeneous graph representation of the healthcare cloud ecosystem, revealing the details of the complex relationships between diverse entities such as users, devices, and digital assets.

Input into this model will be heterogeneous graph representation, which can be denoted as , where denotes the set of nodes, represents the set of edges, defines the node types, and specifies edge types. Each node corresponds to a distinct entity within the system—such as a user (doctor, nurse, administEach node in E may represent a different entity in the system-a user (doctor, nurse, administrator), a device (computer workstation, IoT sensor), or a file (patient record or encrypted document). And edges will be used to indicate the various relationships or interactions happening between any of the above-introduced entities: data transmission, login events, access requests, and device communication.

Each edge is associated with temporal and contextual attributes, such as access time, frequency, and action type. The adjacency matrix dynamically evolves as new events occur, allowing the model to maintain awareness of real-time network activity.

Before feeding the network, node-level and edge-level features are extracted to describe behavioral patterns. For every node , a feature vector is defined as:

17

where each feature represents a measurable behavior:: frequency of data access or retrieval actions,: encoded privilege level or department role,: temporal patterns in login or file activity,: type and trust score of the accessing device,: network origin or regional identifier.

Edges are also described by contextual vectors encapsulating access duration, data volume, and request type. These enriched features enable the model to learn distinct relational patterns, thereby aiding the differentiation between legitimate and abnormal behavior sequences.

Unlike the conventional graph neural networks treating any and all nodes and edges as equivalent, the MGNN accepts the heterogeneity by treating each type of relation separately and combining them through a meta-path-based attention mechanism. Let Φ be the meta-path denoting a specific relationship chain (e.g., User → Device → File). For each meta-path , the embedding of node is calculated as:

18

where denotes the attention weight between nodes and under the meta-path , is the transformation matrix for that path, and is a nonlinear activation function.This operation is repeated for all relation types, after which the embeddings are fused to form a unified mixed representation:

19

Here, is a learned importance coefficient that determines the contribution of each relational pattern to the final embedding. This mechanism enables the MGNN to recognize both explicit and implicit dependencies among heterogeneous entities — for instance, a pattern where a non-medical staff member repeatedly accesses restricted patient records through a shared workstation.

Once the embeddings are obtained, they are passed through a linear classification layer that evaluates behavioral normality. The decision boundary is learned from labeled historical access logs, distinguishing legitimate patterns from anomalies. For each event representation , the output prediction is defined as:

20

where and are learnable parameters of the classification layer. Events classified as anomalous are flagged for further inspection or automatically trigger an alert.

The model detects threats through deviation-based learning, comparing real-time embeddings with expected behavioral clusters. When deviations exceed a defined threshold , the event is marked suspicious:

21

The anomaly decision threshold δ is not arbitrarily fixed but is data-adaptive and statistically grounded. During the training phase, embeddings corresponding to normal behavior are used to estimate the centroid and dispersion of the latent embedding space. Specifically, δ is determined from the empirical distribution of Euclidean distances computed over validation samples containing only legitimate access patterns. The threshold is set as a high-percentile boundary of this distribution to balance sensitivity and false alarm rates, defined as:

22

where and denote the mean and standard deviation of normal-distance scores, and λ is a tunable scaling factor. In the experimental evaluation, λ was set to 3.0, corresponding to a 99.7% confidence interval under the Gaussian assumption, which effectively filters rare but legitimate deviations while preserving high recall for anomalous behaviors. This configuration resulted in an optimal trade-off between detection accuracy and false-positive rate and was consistently applied across all experiments, yielding stable performance under varying workload and user-density conditions. The adaptive nature of δ allows the MGNN to remain robust against evolving access patterns and concept drift in dynamic cloud healthcare environments. To detect both existing and emerging threats, such as insider abuse, privilege escalation, and anomalous access sequences, the MGNN was trained with simulated access log files and anonymized healthcare transaction datasets. All log entries contain time stamps, user identifiers, device metadata, and file access information. The dataset used for training was a combination of normal operational activities and syntactically generated abnormal patterns that imitated insider and external attack patterns. Through heterogeneous graph modeling, the MGNN captures complex relational dependencies of various data types/user interactions. It dynamically changes the behavioral contexts for real-time anomalous activity detection across cloud infrastructure when learning meta-path-specific embeddings and attention weights. This module is integrated into the overall flow to ensure threat monitoring and intelligent threat detection beyond any performance impact on the system, as shown in Fig. 4.

Fig. 4.

Overall architecture of the multi-graph neural network (MGNN)–based fusion model.

Result and discussion

This section presents a detailed performance analysis of the proposed multi-layered cloud security framework and compares it with existing systems, namely, SEHR-BC-CHG (16), DKS-CWH (17), RBGTFN (24), and AHO-ABE³¹. Experiments were conducted to analyze the performance of the model from various perspectives, including efficiency, reliability, and adaptability under different computational-operational conditions. The comparative assessment considers dataset characteristics or experimental configurations other than different performance evaluation criteria, such as accuracy, loss, execution time, throughput, energy utilization, computation time, latency, and memory consumption. Each subsection is dedicated to a discussion of the comparative performance of the proposed system against the baselines along parameters such as data protection increment, faster response time, and improved resource utilization. Furthermore, the discussion interprets the results regarding the collaboration of integrated encryption, key optimization, and anomaly detection mechanisms to provide secure, scalable, and high-performance management of healthcare data in cloud environments.

Dataset description

This dataset was collected from the Electronic Health Record (EHR) repository on Kaggle, provided by Saurabh Shahane. The database consists of actual patient records retrieved from a private healthcare establishment in Indonesia and intends to predict the outcome of health treatment services. Each record corresponds to an examination of the patient in the laboratory which has hematology parameters, such as Haematocrit, Haemoglobins, Erythrocyte count, Leucocyte level, Thrombocyte, Mean Corpuscular Hemoglobin (MCH), Mean Corpuscular Hemoglobin Concentration (MCHC), Mean Corpuscular Volume (MCV), among others. To provide a complete setting for individual differences in clinical outcomes, the anthropomorphic data also include Age and Sex. The target attribute of SOURCE is a categorical variable classified into two categories, namely “in-care” or “out-care” which indicates the evidence for being confined under the care of the hospital or discharge. Thus, the structure of this dataset is suitable for a supervised classification framework that is devised to predict later treatment categories based on laboratory findings.

Experimental configuration

The proposed multilayer secure cloud framework, which amalgamates SR-ABE, HO-based key generation, and SHA-256 integrity verification with MGNN anomaly detection, is experimentally evaluated in a controlled computing environment to ensure stable and repeatable results. All modules, including the cryptographic, integrity verification, and neural network components, were implemented in Python 3.10. Deep learning and graph modeling were accomplished using PyTorch 2.2 and PyTorch Geometric, respectively, while data analysis and numerical computation were handled using NumPy 1.26 and Pandas 2.2. The results were visualized, and encryption and decryption performance graphs were created using Matplotlib 3.8 and Seaborn 0.13. For secure hashing and random key generation, Hashlib and random from the OS library were used, whereas Scikit-learn 1.5 was used to evaluate the classification metrics.

The hardware configuration therefore consisted of a 64-bit Ubuntu 22.04 LTS operating system powered by an Intel Core i7-10750H CPU (2.60 GHz, 6 cores), 16 GB DDR4 RAM, and an NVIDIA GeForce RTX 2060 GPU with 6 GB VRAM. Essentially, this setup is sufficient for performing encryption operations, model training, and real-time anomaly detection without facing memory or latency bottlenecks. A realistic experimental setting, the eHealth cloud, with 200 data users, one key authority, and one semi-trusted cloud server, was configured. The cryptographic parameters included kinetic keys of 128, 192, and 256 bits, and the performance scalability was measured on data blocks. The average values of encryption and decryption were analyzed for different key configurations, whereas throughput and latency were evaluated under concurrent access conditions for a feasibility work in the field.

Experimental results

The analysis shown in Fig. 5 implies that the proposed framework effectively attains stability in learning and accurately predicts results, through which it surpasses every other existing model, such as RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. The graphs for training and testing portray a consistent increase in accuracy alongside a decrease in loss, which indicates good convergence and pre-empts overfitting during training. This model obtained 99.1% training accuracy and 96.7% testing accuracy, far surpassing all baseline methods and showing great generalizability, coupled with training and test losses minimized to 0.0126 and 0.0357, respectively. Such a figure of merit shows that given any representation of the data, it is learned with all but the smallest amount of error. Greater rising slopes in accuracy, together with greater edges in loss downward, show the security efficiency and computational robustness of the entire system under the influence of Haze Optimization-desired key management and the MGNN-based adaptive learning module, as shown in Fig. 5.

Fig. 5.

Comparative analysis of training and testing accuracy and loss between the proposed model and existing methods.

Execution time analysis

The performance of execution, as depicted in Fig. 6, shows a more detailed analysis of the behavior of the proposed system with respect to different computational loads with respect to the number of training epochs and the number of active users. As shown, the proposed framework consistently demonstrated lower execution times under all conditions than the baseline models, namely RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. For example, at an epoch of 50, the proposed system has an average execution time of 10.4 ms, whereas the competing models take longer than 13–15 ms. Similarly, as the user count is increased from 20 to 220, the proposed model shows an uptrend of execution time with moderate steadiness—an increase from 20 to 60 ms, which is, however, much lower than the 77–108 ms observed by other methods.

Fig. 6.

Comparative evaluation of execution time with respect to the number of epochs and user count across different models.

Throughput evaluation

The throughput analysis shown in Fig. 7 demonstrates the optimum communication efficiency with the data handling capabilities of the proposed secure healthcare framework. As shown, the proposed model outperforms the RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE in terms of throughput through varying epochs and user counts. When the epochs were increased, there remained a constant throughput of 901 kbps at 10 epochs, and at higher epochs, maintained 852 kbps even at 50 epochs. Meanwhile, the competing models witnessed a dramatic fall, which was attributable to computational overheads and inefficient encryption cycles. In addition, under user scaling, the proposed framework would show superlative stability by obtaining 990 kbps at 20 users and recovering 700 kbps at 220 users compared to all the baseline techniques.

Fig. 7.

Throughput comparison of the proposed model and existing schemes across epochs and varying user counts.

Energy consumption analysis

The energy utilization analysis presented in Fig. 8 indicates that the proposed encryption and optimization framework exhibits a more efficient energy consumption profile than the baseline models RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. With an increase in epochs, the proposed framework shows almost negligible power consumption, beginning with 0.099 J at 10 epochs and ending at only 0.214 J at the 50th. Existing approaches tend to consume higher energy of up to 0.28 J or even more because of prolonged encryption cycles and redundant key computations. Furthermore, across varying user numbers, the proposed framework maintains an optimal energy consumption growth pattern, starting from 0.075 J for 20 users to 0.205 J for 220 users, whereas others show a steep rise in energy costs.

Fig. 8.

Energy consumption comparison across various models over epochs and user scalability.

Computational time evaluation

The relative evaluation of the computation times, as illustrated in Fig. 9, offers a comprehensive comparison of the performance of the proposed system with that of existing encryption-based health models. As shown, the proposed method invariably exhibits a lower computation time as the number of epochs and users increases. The model, in particular, gives a minimum computation time of 0.098 s at 10 epochs and only 1.74 s at 50 epochs, making it better than others such as RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE as they demonstrate a slower convergence rate owing to their heavier key management and encryption operations. Even under user scalability testing, the proposed framework is superior, requiring only 0.38 s at 20 users and scaling impeccably to 2.3 s at 220 users.

Fig. 9.

Computation time comparison of different models across epochs and user counts.

Latency performance analysis

The latency evaluation results in Fig. 10 show that this framework is sensitive and efficient in terms of data transmission under varying operational modes. The results show that the proposed system latency performance is statistically significantly lower than that of the RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE traditional models at growing epochs and increasing user bases. The proposed framework has the lowest recorded latency of 3.3 ms at 10 epochs and just 6.6 ms at 50 epochs, while the previous schemes have a latency period of approximately 7.2 and 8.8 ms, predominantly because of the higher encryption complexity and subsequent processing of data. Likewise, with an increasing number of active users from 20 to 220, the proposed method’s latency range is maintained between 2.9 and 6.1 ms, exhibiting a linear gradual lift, in contrast to the other frameworks that confront an exponential rise beyond 8 ms owing to processing bottlenecks, as shown in Fig. 10.

Fig. 10.

Latency comparison of proposed and existing models across epochs and user scalability.

Memory utilization analysis

As shown in Fig. 11, the performance of the memory is assessed to clearly indicate that the given architecture offers superior memory efficiency compared to traditional architectural models such as RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. Overall, the proposed system exhibits considerably lower memory consumption, which is a testament to its lightweight design and optimized cryptographic processing. Memory consumption is only 2.45 GB after 10 epochs in the proposed model, which is less than RBGTFN and SEHR-BC-CHG that require 3.30 and 3.15 GB, respectively. Even after 50 epochs of training, the proposed method only increased to 5.00 GB, while all the others increased to above 6.20 GB owing to inefficient key management techniques and redundant operations in storing data. Similarly, when subjected to different user loads between 20 and 220, the proposed system scaled effectively, with memory utilization increasing from 2.2 GB to 5.7 GB. Baseline techniques yielded much higher numbers for this type of testing.

Fig. 11.

Memory utilization comparison of proposed and existing models across epochs and user scalability.

Security performance evaluation

To quantitatively evaluate the security effectiveness of different encryption models, this work adopts Data Confidentiality Rate (DCR) and Data Integrity Rate (DIR) as core performance metrics. The Data Confidentiality Rate measures the system”s ability to prevent unauthorized entities from successfully accessing protected medical records. It is computed as the ratio of denied unauthorized access attempts to the total number of unauthorized access requests issued during the evaluation period, expressed as a percentage:

23

where denotes the number of access attempts correctly rejected by the encryption and access-control mechanisms, and represents the total unauthorized access attempts generated in the experimental workload. A higher DCR indicates stronger resistance against privilege misuse, key compromise, and collusion-based attacks.

The Data Integrity Rate quantifies the system”s capability to detect and prevent unauthorized modification of encrypted medical data during storage or transmission. It is defined as the percentage of data objects whose integrity status is correctly verified without mismatch under intentional tampering and normal access conditions:

24

where denotes the number of data items that successfully pass cryptographic hash verification (SHA-256) without alteration, and is the total number of integrity validation operations performed. This metric reflects the robustness of the integrity verification mechanism in detecting silent data corruption, replay attacks, and in-transit manipulation. Both metrics are evaluated under identical experimental conditions using the same dataset and attack simulation parameters to ensure fairness and comparability across encryption models.

The comparative security evaluation, summarized in Table 2, shows the strength of the proposed framework in upholding high standards of data confidentiality and integrity against all conventional models, such as RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. The proposed system achieved a confidentiality rate of 99.12%, which is much higher than that of any other existing techniques (between 94 and 98%), and the system had an integrity success rate of 98.53%.

Table 2.

Comparison of data confidentiality and integrity among various encryption models.

Techniques	Data confidentiality (%)	Data integrity (%)
RBGTFN	95.29	94.13
SEHR-BC-CHG	96.26	95.09
DKS-CWH	98.20	97.01
AHO-ABE	97.23	96.05
Proposed	99.12	98.53

In public and private key operations, the cryptographic performance evaluation presented in Table 3 assesses the efficiency of the proposed framework. In all the measured parameters, including public key generation, public encoding, public decoding, private key generation, private encoding, and private decoding, the proposed approach exhibited the best performance with minimum time in comparison to any of the seven previously mentioned traditional models: RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. Public key generation is recorded at only 1600 ms and public encoding at 1250 ms with the proposed system, while the other two, DKS-CWH and AHO-ABE, take approximately 1800-2000ms. The private key operations demonstrate a 2100 ms key generation and a 1500 ms decoding with the proposed framework, which is also faster than other frameworks, whose running time exceeds 2500 ms owing to redundant encryption cycles and inefficient key structures.

Table 3.

Comparative analysis of cryptographic computation efficiency across different encryption models.

Methods	Public key generation (ms)	Public encoding (ms)	Public decoding (ms)	Private key generation (ms)	Private encoding (ms)	Private decoding (ms)
RBGTFN	2200	1650	880	2750	2750	1980
SEHR-BC-CHG	2100	1575	840	2625	2625	1890
DKS-CWH	1800	1350	720	2250	2250	1620
AHO-ABE	2000	1500	800	2500	2500	1800
Proposed	1600	1250	650	2100	2000	1500

Discussion

Practical applicability and generalizability

Although the proposed framework was implemented and evaluated in a controlled experimental environment, the obtained results provide strong evidence of its practical feasibility in real-world cloud-based healthcare settings. The measured low access latency primarily reflects the computational overhead introduced by encryption, integrity verification, and MGNN-based monitoring, which are architecture-intrinsic properties that are largely independent of specific clinical workflows. Because the framework is designed to offload computationally intensive operations to cloud or edge servers while limiting end-device tasks to lightweight cryptographic primitives, similar low-latency behavior is expected under realistic deployment conditions, subject to network and infrastructure variability. Furthermore, the experimental evaluation employs a real-world electronic health record (EHR) dataset collected from Indonesia, which contains heterogeneous user roles, access patterns, and clinical interactions that are representative of operational healthcare environments. Although only a single dataset was used, the proposed MGNN-based modeling relies on structural and relational interaction patterns rather than dataset-specific features, enabling the learned representations to generalize across different healthcare systems with comparable access control and data-sharing behaviors. Therefore, the observed performance indicates the promising generalization capability of the proposed model, although broader cross-dataset and cross-institution validation remains an important direction for future work.

Deployment feasibility on resource-constrained devices

The execution time and energy consumption metrics were obtained using a desktop-class CPU (Intel i7-10750H) and GPU (NVIDIA RTX 2060) to ensure a controlled and reproducible benchmarking. These measurements should be interpreted as upper-bound system-level indicators rather than direct representations of the performance of resource-constrained medical devices. In the proposed architecture, computationally intensive operations, such as SR-ABE partial decryption, Haze Optimization–based key refinement, and MGNN-based anomaly inference, are explicitly offloaded to cloud or edge servers, whereas embedded medical devices perform only lightweight tasks, including symmetric encryption/decryption, SHA-256 hashing, and metadata transmission. This architectural partitioning significantly reduces computation, energy consumption, and latency at the device level and enables scalable deployment across heterogeneous medical IoT environments without imposing prohibitive resource requirements on the devices.

Conclusion

This work aimed to secure cloud-based e-health systems not with a single control but with a pipeline of mutually reinforcing controls, and the experiments confirmed that such layering is necessary. At the entry point of the system, health data are encrypted with Server-Aided Revocable Attribute-Based Encryption, meaning that access is tied to attributes (role, department, duty), and the cloud is allowed to help with decryption without ever seeing the plaintext; this is crucial for hospitals that must outsource storage but cannot outsource trust. To stop weak or guessable keys from becoming the weakest link, the SR-ABE layer is supplied with keys produced by Haze Optimization, which searches the key space for high-entropy candidates and selects the one with the best fitness; in practice, this kept the key-management time lower than that of RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE, while still increasing resistance to key compromise. Every encrypted object is then paired with a SHA-256 hash so that when the record is later pulled from the cloud, the framework recomputes the digest and rejects any file whose hash is different. This is a simple step, but the results show that it preserved integrity in 98.53% of the cases, which is higher than all the comparison methods. The model reached 99.1% training accuracy and 96.7% testing accuracy, with very low losses (0.0126 train, 0.0357 test), which means that it could distinguish normal medical access from suspicious behavior with high confidence. Performance experiments further showed that the framework is not only secure but also fast: execution time dropped to 10.4 ms at 50 epochs, throughput stayed high at 852–901 kbps, energy stayed low at 0.214 J, latency stayed within 3.3–6.6 ms, memory stayed around 2.45–5.0 MB, and security indicators peaked at 99.12% confidentiality, which is better than RBGTFN, SEHR-BC-CHG, DKS-CWH, and AHO-ABE. Going forward, the same architecture can be extended in three directions: (i) plug the integrity hashes into a permissioned blockchain so that tampering is provably logged across institutions; (ii) fine-tune the MGNN on streaming hospital logs so that the detector adapts to seasonal and staff-shift changes without retraining from scratch; and (iii) offload SR-ABE and HO computations to edge gateways near medical IoT devices to reduce latency even further in emergency or tele-ICU scenarios.

Limitations and future scope

The proposed framework was experimentally evaluated using a real-world electronic health record (EHR) dataset comprising structured hematology parameters from Indonesian patients, which provided a consistent and realistic basis for validating access control, integrity assurance, and behavioral anomaly detection. While this dataset captures heterogeneous user roles and access patterns typical of operational healthcare systems, it does not encompass unstructured imaging modalities (e.g., X-ray, CT, or MRI) or multimodal data combinations such as genomic–clinical records. This limitation is primarily due to the focus of the current work on security architecture and access behavior modeling rather than data-type-specific analytics. Comprehensive validation across diverse data modalities and multi-institutional datasets is an important direction for future work to further substantiate cross-domain robustness.

Author contributions

Conceptualization: Muthuvel. S Data curation: Muthuvel. S Investigation: Priya. S, Sampath Kumar. K Methodology: Muthuvel. S Project administration: Priya. S, Sampath Kumar. K Supervision: Priya. S, Sampath Kumar. K Validation: Priya. S, Sampath Kumar. K Writing—original draft: Muthuvel. S

Funding

This research received no external funding.

Data availability

All data analyzed during this work are included in this article.

Declarations

Competing interests

The authors declare no competing interests.