A federated edge intelligence framework with trust based access control for secure and privacy preserving IoT systems

Abstract

The rapid growth of Internet of Things (IoT) ecosystems has generated substantial industrial progress, yet it has also introduced intricate security and privacy issues. IoT deployments cannot be properly supported with traditional cloud-centric approaches because they require improved bandwidth utilization, reduced latency, and enhanced trust mechanisms. The research proposes Artificial Intelligence-Driven Secure Edge Trust Framework (AI-SET), which establishes a comprehensive edge-based security design that connects network intrusion detection with federated learning capabilities to implement adaptive trust-based access control for IoT system protection. The AI-SET framework comprises three central elements. Real-time anomaly detection at the network edge through the Edge-Resident Intrusion Detection System operates with lightweight AI algorithms to minimize dependency on centralized systems. Privacy-preserving federated learning utilizes the modified FedAvg algorithm, which is supported by differential privacy and homomorphic encryption. Security measures enabled by this model allow algorithms to be trained across decentralized sources that contain heterogeneous and non-identically distributed (non-IID) data. A dynamic access control system utilizes trust assessment models to evaluate device context and behavior for real-time permission evaluations. The framework undergoes validation by running tests with the NAB dataset, supported by Jetson Nano and Raspberry Pi edge devices, and tools including Suricata, Metasploit, and the WAZUH threat platform. Evidence shows that AI-SET boasts higher accuracy in intrusion detection, enhanced communication performance, and superior access control security compared to standard approaches. AI-SET demonstrates immunity against attempted model poisoning attacks and unauthorized system breaches, achieving this protection while maintaining low operational costs and ensuring secure data privacy. The research presents AI-SET as an adaptable, resilient, and sensitive-minded security framework for future IoT systems, through its holistic control of edge intelligence, secure network operations, and automated trust management.

Introduction

Internet of Things growth patterns have established intelligent spaces where devices, sensors, and actuators operate independently to share information and react to current events. The increasing number of connected devices has driven the digital transformation of key industries, including healthcare, manufacturing, transportation, and energy. The wide array of cyber threats targeting IoT systems stems from their exceptional interconnectivity and diverse device nature, as they face threats such as data breaches, denial-of-service attacks, spoofing, and unauthorized access¹. The growing volume of data, coupled with the increasing number of connected devices, makes it essential to develop effective security measures that work efficiently and scale intelligently. Financial sector security systems primarily rely on central cloud infrastructure systems to collect and analyze information. Although such methods sometimes prove effective, they do not meet the requirements of current IoT systems, which demand low-latency performance alongside constrained bandwidth usage. Centralized data server transmission exposes sensitive data to major privacy threats and puts information at risk of compromise or interception. The emergence of edge computing brings computational capabilities closer to data points, delivering low-latency processing capabilities and providing context awareness alongside bandwidth efficiency. The promise of edge computing comes with various challenges that affect trust maintenance, the implementation of dynamic access policies, and secure inter-device collaboration among resource-limited systems. According to², the lack of standardized edge security frameworks increases the risk of compromised devices and unreliable data exchange. Relevant to these urgent IT challenges, this paper introduces AI-SET, which utilizes AI to create a unified framework that addresses the security, long-term, privacy, and resilience needs of contemporary IoT platforms through three integrated approaches.

Edge-resident intrusion detection system

A lightweight anomaly detection model deployed at the network edge to provide real-time, localized threat detection without dependency on centralized cloud services.

• Privacy-preserving federated learning:

• A customized Federated Averaging (FedAvg) algorithm integrated with differential privacy and homomorphic encryption enables secure, decentralized model training across non-IID IoT devices without exposing raw data.

• AI-driven trust-based access control:

• A dynamic, context-aware trust evaluation system that continuously adapts access permissions based on device behavior, enhancing the resilience of IoT systems against evolving threats.

The AI-SET framework strengthens IoT security and promotes scalability, interoperability, and user trust by reducing attack surfaces and enabling secure edge collaboration. Through extensive experimentation under realistic attack scenarios, AI-SET demonstrates superior performance in anomaly detection accuracy, privacy preservation, access control adaptability, and system robustness.

This paper lays the groundwork for the next generation of intelligent, secure, and autonomous IoT environments where security is not a reactive mechanism but an embedded, adaptive, and proactive feature of the system architecture³. The following chapters provide an in-depth discussion of the design, implementation, evaluation, and implications of the AI-SET framework, illustrating its potential to transform security within edge-enabled IoT environments.

Research gap

Despite significant advancements in IoT security, notable research gaps hinder the widespread deployment of intelligent, scalable, and privacy-preserving security solutions in edge-enabled environments. Existing intrusion detection systems (IDSs) for IoT rely on resource-intensive cloud-based infrastructures or have limited detection capabilities when deployed on constrained edge devices⁴. These approaches often fail to deliver real-time responsiveness and are prone to high latency, which is unacceptable in critical IoT applications. Moreover, while Federated Learning (FL) has emerged as a promising solution for decentralized model training, its integration into IoT environments remains nascent. Current implementations inadequately address the challenges of device heterogeneity, non-IID data distribution, and privacy leakage through gradient sharing. Most solutions lack robust encryption mechanisms or fail to incorporate differential privacy safeguards, leaving systems vulnerable to data inference attacks.

Another major gap lies in static access control systems that do not account for the dynamic and context-aware nature of IoT device behavior. Existing models typically assign permissions based on predefined rules without continuously adapting to changes in trustworthiness or behavior anomalies, making them ineffective against evolving threats and insider attacks.

To bridge these gaps, the proposed AI-SET framework uniquely combines:

• Lightweight edge-resident anomaly detection,

• Federated learning enhanced with privacy-preserving mechanisms, and

• Real-time, trust-based adaptive access control.

AI-SET significantly advances the state of the art in securing edge-enabled IoT ecosystems by addressing latency, scalability, privacy, and trust together within a unified architecture.

Problem definition and research objectives

Despite the rapid growth of IoT environments, significant challenges remain in achieving real-time security, preserving data privacy during collaborative learning, and implementing adaptive access control on resource-constrained edge devices. Many existing approaches rely on centralized architectures, which result in high latency and expose sensitive data to potential privacy breaches. Moreover, current solutions tend to address these issues independently rather than in an integrated manner.

To address these gaps, our study aims to:

• Develop a lightweight Intrusion Detection System: Deployable at the edge for real-time anomaly detection.

• Implement Privacy-Preserving Federated Learning: Leveraging Differential Privacy and Homomorphic Encryption to Securely Aggregate Models from Non-IID IoT Devices.

• Design a Trust-Based Dynamic Access Control Mechanism: Enabling adaptive permission settings based on the continuous evaluation of device behavior and contextual factors.

These objectives are realized within the proposed AI-SET framework, which integrates edge intelligence, secure federated learning, and dynamic trust evaluation into one cohesive IoT security architecture.

Related work

Recent advancements in IoT security have explored multiple paradigms, such as edge computing, federated learning, anomaly detection, and AI-driven access control. However, a holistic integration of these technologies remains limited. Introduced an AI-based intrusion detection system for edge computing environments using federated learning. Their model achieved real-time detection but was limited by its constrained attack coverage and model interpretability. An end-to-end healthcare IoT security framework leveraging blockchain and encryption, but lacked adaptability in dynamic environments and device scalability. An edge gateway framework that combines AI with cybersecurity provides enhanced real-time capabilities. The system primarily addressed network-level attacks but lacked capabilities for federated learning and trust-based control. To research federated edge learning, which demonstrated that secure aggregation methods are essential because existing privacy and scalability issues continue to exist.

Introduced the significance of AI for IoT security through anomaly detection models alongside behavioral analysis, while omitting discussions on edge-device design limitations. The research investigates self-governing federated learning detection of network intrusions within public networks; however, it does not incorporate access control mechanisms or profile device activities. AI implementations in IoT edge security, however, their work did not include dynamic access control and distributed node collaborative intelligence. The literature reveals a significant interest in tracebacks with AI understanding for mobile edge systems, although the research currently focuses on packet analysis rather than comprehensive IoT security solutions. The existing solutions work independently to handle the detected problems. Lightweight anomaly detection, privacy-preserving learning, and trust-based adaptive access control at the edge require a single, integrated solution that current research lacks. Researchers focus on three solutions that improve IoT security in the reviewed literature: federated learning, blockchain, and edge AI. The research studies^5–7 demonstrate how blockchain-enabled FL boosts data privacy, but they do not provide adequate solutions for universal implementation at edge locations. The examined studies^8,9 explore communication optimization and non-IID data processing without integrating dynamic access control systems or anomaly detection capabilities. The surveys detailed in¹⁰ and the implementations presented in¹¹ focus on individual security aspects of IoT, yet fail to combine trust-based privacy mechanisms and learning techniques into a unified approach. Table 1 describes the summary of the literature survey.

Table 1.

Summary of related works on IoT security approaches.

Reference	Focus area	Approach	Limitations
Lu et al.5	Federated learning for industrial IoT	Blockchain-integrated FL for privacy	Limited attack diversity and interpretability
Kumar et al.6	Secure FL in consumer IoT	Blockchain + encryption-based FL	Poor adaptability in dynamic IoT environments
Issa et al.7	Federated learning & blockchain survey	A comprehensive review of FL + blockchain	Lacks unified edge deployment models
Wang et al.8	Clustered FL on non-IID data	FedAvg with clustering	No integration with access control or anomaly detection
Mills et al.9	FL efficiency in wireless edge	Comm-efficient FL for IoT	Limited trust and privacy mechanisms
Khoei et al.10	DL challenges in IoT	Survey of DL models for Edge devices	No practical implementation or access control focus
Thakur et al.11	Secure smart homes	Edge-AI for home IoT	Lack of trust-based access control and FL integration
Zhang et al.20	Federated learning in healthcare IoT	Lightweight FL with local privacy mechanisms	No trust-based access control or real-time anomaly detection
Ahmed et al.21	Fog-based IoT access control	Policy-aware fog security framework	Lacks federated training and dynamic behavioral analysis
Lin et al.22	Edge AI with blockchain	Secure edge framework with encryption layers	No evaluation with heterogeneous edge environments or FL

To address the security, privacy, and access control issues presented in the following sections, the present study proposes the AI-SET framework. In the next section, the technical architecture and design of AI-SET are presented, along with an outline of its main components and their collaboration in protecting edge-enabled systems between IoT devices.

Even recent works in 2024 have utilized federated learning and access control as part of developing the IoT security domain. A lightweight federated IoT framework for healthcare that achieved privacy enhancements but lacked dynamic trust modeling²⁰. An effective access control mechanism that suits fog-IoT systems, but it lacks the incorporation of federated learning and real-time anomaly detection²¹. A study on blockchain and edge AI secure architecture, which had not been tested on realistic non-IID datasets yet²². These undertakings demonstrate the growing need for edge security, which nonetheless only addresses fragmented aspects of an integrated defense system in IoT. What makes AI-SET unique is the combination of lightweight anomaly detection, privacy-preserving model training, and adaptive trust-based access control into one cohesive unit, optimized to run in edge environments with limited resources.

Proposed framework: AI-SET

We begin by discussing the first component of AI-SET: the Edge-Resident Intrusion Detection System, which enables real-time anomaly detection at the network edge.

Edge-resident intrusion detection system

The Edge-Resident Intrusion Detection System operates as the initial essential component of the AI-SET framework. The system functions to detect anomalies in real-time directly at the edge devices, resulting in decreased data transfers to the central cloud server. Implementing this system reduces latency while maintaining data privacy and enabling bandwidth to function more efficiently. The detection system learns normal behavioral patterns from data streams while identifying suspicious activities through deviations in the data. The detection mechanism installed near or inside IoT devices through IDoS ensures threat analysis with low-latency performance near the devices.

Key features include:

• Low resource consumption: Designed for ARM-based edge devices like Raspberry Pi and Jetson Nano.

• The system detects anomalies instantly when they occur, thus enabling rapid threat response.

• The modular design structure allows the deployment of this system across various device environments.

IDoS obtains training from real-world NAB datasets by combining them with open-source security tools, such as Suricata and WAZUH, for traffic analysis and signature rule development⁹. When combined with a hybrid detection system, a security system becomes more resistant to identifying zero-day attacks.

However, unlike traditional IDS models, the newly proposed Edge-Resident Intrusion Detection System has several innovations specific to IoT edge settings. First, it utilizes a hybrid detection mechanism that integrates rule-based alerts (via Suricata) and adaptive AI forecasts, enabling context-aware threat responses. Second, we deploy a model-adaptive resource controller that enables the adaptive reduction of LSTM depth and the dexterity of a decision tree with the available edge-node resources (CPU and memory). Third, the LSTM model is optimized using truncated backpropagation and quantized weight layers to reduce energy consumption while maintaining temporal detection fidelity. The system also features a multi-modal set of capabilities, including time-series metadata and machine-specific behavioral indicators, which are more sensitive to stealthy or zero-day attacks that evade standard rule-based filters. These are finely tuned improvements that overcome the computational and environmental limitations inherent in real-world edge computing deployments for IoT systems.

Privacy-preserving federated learning at the edge

AI-SET implements a federated averaging-based decentralized collaborative learning architecture as its second essential component. This component enables IoT devices to train a unified model while maintaining their privacy, thereby minimizing bandwidth usage during training operations. A security mechanism for privacy protection has been developed for federated learning by combining three elements. Differential privacy protects system security by adding noise to updates, preventing attackers from making inferences. Homomorphic Encryption: Secures the transmission of model parameters during aggregation. The combined protection mechanism enables nodes to maintain local possession of raw sensory information while allowing them to participate in universal model training.

Key capabilities include:

The system handles data distributions that differ from case to case between devices.

• The system maintains accurate performance even during periods of restricted network communication.

• Resilience against model poisoning and inference attacks.

• A secure server using encrypted procedures aggregates and updates the federated learning system. The experimental testing indicates that the system builds upon FL setups to deliver better model results and minimizes network traffic compared to standard FL deployments.

The Gaussian differential privacy is set to epsilon = 1.0 and delta = 1e-5, where noise is added to the gradients before they are summed. In the case of homomorphic encryption, we will use CKKS and TenSEAL with 128-bit security. The lightweight symmetric key exchange communicates encryption keys on a per-session basis, and encryption parameters are optimized with computation overhead in mind, without compromising model performance.

Trust-based dynamic access control mechanism

AI-SET integrates an AI-driven access control system that implements trust-based security measures to apply dynamic context-aware policies. This security system distinguishes itself from static ACLs by performing ongoing behavioral analysis of IoT devices through time-stamped records and real-time evaluations to determine trust levels.

Trust scores are computed using:

• Anomaly signals from IDoS,

• Contextual factors (e.g., time of access, location, behavior history), and

A system utilizes machine learning algorithms for detecting irregular activities from ordinary system behavior patterns. The system adjusts access permissions based on these acquired trust scores. The trust rating of IoT devices determines whether they gain unrestricted access or are restricted to specific environments. Figure 1 shows the simplified AI-SET Framework. Real-time trust evaluation and decision-making. The system minimizes attacks from insiders or breaches of secure network nodes. The system can handle IoT environments comprising multiple devices with varying behavioral patterns. Studies through simulation have revealed that the trust-based method reduces unauthorized access attempts and enhances security measures against compromised nodes, while maintaining operational integrity for genuine devices. Figure 2 demonstrates the detailed AI-SET framework.

Fig. 1.

Simplified framework.

Fig. 2.

An artificial intelligence-driven secure edge trust framework.

Hierarchical weights of the trust scores (α, β, γ) were set a priori, based on heuristic choices that gave paramount importance to anomaly detection; however, they were later optimized using a sensitivity test. We reduced and increased the parameters step by step and examined the accuracy of access control for each test case to determine the optimal setting. The resulting final values yielded consistent results, with a high block rate of unauthorized access and low false positives.

A system of mathematical equations describes the fundamental computational processes for the three main subsystems of the AI-SET framework.

1. Data preprocessing

   Standardization of input feature:

   

   where x: original input, μ: mean, σ: standard deviation.
2. Anomaly detection

   • Anomaly score (mean absolute error):′

     
   • Alert logic:

     
   • Binary classification output:

     
   • Loss function:

     
   • Gradient calculation:

     
3. Federated learning

   • Gradient perturbation using differential privacy:

     
   • Encryption invariant in homomorphic encryption:

     
   • Federated averaging:

     
   • Model update using SGD:

     
4. Trust-based access control

   • Trust score formula:

     

     Where H_t: historical behavior score, A_t: anomaly score, C_t: contextual trust factor.
   • Historical behavior score:

     
   • Dynamic access threshold:

     
   • Access permission decision:

     
5. Evaluation metrics

   • Accuracy:

     
   • Precision:

     
   • Recall:

     

6. Communication overhead

   Encrypted model update size:

   
   Communication rate:

   

These mathematical expressions form the fundamental basis for the AI-SET algorithm to achieve its objectives of supporting IoT system privacy, efficiency, and security. A new multi-phase AI-SET algorithm operationalizes the integration between edge-based intelligence federated learning and trust-aware access control in IoT environments. A single secure and efficient edge-computing architecture results from merging three subsystems through this algorithm.

AI-SET algorithm: secure edge trust framework for IoT

Step-by-Step Workflow

1. Data acquisition & preprocessing

   • The flow of real-time IoT data occurs through devices to edge nodes.
   • The edge nodes apply two functions to data: they normalize and extract features from raw information streams while also performing filtering activities.
2. Edge-resident intrusion detection

   • Data analysis occurs locally by executing lightweight ML models (including decision trees and LSTM) on the data.
   • The calculation of anomaly score takes place through
   • A suspicious activity occurs when Aₛ exceeds θ.
3. Federated learning with privacy

   • The edge nodes execute individual model training processes using their locally collected data.
   • The application implements differential privacy to protect gradients by adding Gaussian noise through the formula Δw* = Δw + (0, σ²).
   • The system applies Homomorphic Encryption to protect Δw as E(Δw*) through encryption.
   • The FedAvg method executes at the aggregator through this operation:
4. Dynamic trust-based access control

   • The trust score formula utilizes a weighted sum of historical trust (H) and activity assessment (A), along with a privacy enhancement parameter (C), to compute T_t.
   • The access decision relies on comparing the token duration T_t and the threshold value τ. Access is offered if T_t surpasses τ, with P_access set to 1, while denial occurs when T_t remains below τ.
   • Sandboxing or denying access constitutes the processing of low-trust devices.

After detailing the design and methodology of AI-SET, we proceed to evaluate its performance through simulation. The next section outlines the experimental environment, datasets, tools, and evaluation metrics used in our analysis.

Software and tools used

The following open-source software tools and libraries were used in this study:

• TensorFlow Federated v0.19.0—for implementing federated learning algorithms—URL: https://www.tensorflow.org/federated

• TenSEAL v0.3.10—for enabling homomorphic encryption of model parameters—URL: https://github.com/OpenMined/TenSEAL

• Suricata v6.0.10—for intrusion detection and network traffic analysis—URL: https://suricata.io/

• Wazuh v4.4.0—for threat intelligence and event correlation—URL: https://wazuh.com/

• Metasploit Framework v6.3.28—for simulating attacks during testing—URL: https://www.metasploit.com/

• Docker v24.0.5—for containerized deployment of modular components—URL: https://www.docker.com/

• Ubuntu 20.04 LTS—as the base OS environment for edge node simulations—URL: https://ubuntu.com/download/desktop

All software used is open source and was configured on local edge testbeds using Jetson Nano and Raspberry Pi devices.

Experimental setup and evaluation

To begin by configuring the simulation setup to replicate realistic IoT conditions.

Simulation environment

Analysis of AI-SET was conducted in a simulation environment that combined various edge nodes, including Raspberry Pi 4 and NVIDIA Jetson Nano systems¹². The simulation required edge servers and virtual machines working with an Ubuntu operating system to mimic a federation-based IoT network. The evaluation used the following software resources together with the specified datasets:

• Datasets: NAB (Numenta Anomaly Benchmark), UNSW-NB15 for synthetic network anomalies

• Tools: Suricata (intrusion detection engine), WAZUH (security information management), Metasploit (attack emulation)

• FL Framework: TensorFlow Federated

• Encryption Library: PySyft and TenSEAL for privacy-preserving operations

Non-IID partitioning represented IoT conditions through simulated Wi-Fi and LTE links that connected edge devices¹³. In a series of experiments, 50 communication rounds were used to evaluate convergence behavior and communication efficiency.

Metrics used

The evaluation metrics for AI-SET involved detection accuracy and false positive rate, precision and recall measures, trust score convergence, and model convergence time.

Detection Accuracy refers to the correct identification rate of anomalies among all monitored data.

• The identification accuracy of false anomalies is known as the False Positive Rate (FPR).

• Precision and Recall—Balance between true positives and false negatives

• The stability of changing trust evaluation measurements as time progresses is known as Trust Score Convergence.

• The model needs to be tested until it reaches a stable accuracy benchmark for each round of federated learning.

• The total amount of transmitted data shows a pattern during rounds with/without encryption.

• The frequency of blocked unauthorized entry attempts determines the effectiveness of access control.

Results and analysis

During the NAB evaluation, AI-SET delivered a detection accuracy of 94.3% and a false positive rate of 3.2%. IDoS delivered responses of 150 ms to activate immediate threat prevention services. With differential privacy implemented in FedAvg, the model achieved 91.5% accuracy after running 28 training rounds. The performance of homomorphic encryption experienced minimal noticeable distortions to computing efficiency. The trust mechanism successfully prevented 87 percent of unauthorized access attempts. Trust score updates were completed within a period of less than 100 ms. This optimization method yielded two benefits: (1) encrypted updates enlarged the data volume by approximately 12%, and (2) the edge-based approach decreased total bandwidth usage by 30% compared to cloud-based learning methods.

We sought to provide robustness by performing a model poisoning simulation, injecting malignant gradient updates at 10%, 20%, and 30% of the involved nodes. AI-SET achieved 85% accuracy in the model under 20% poison, and it gracefully faded after the poisoned levels were reached. Although we have developed a robust trust system that is highly flexible in response to behavioral anomalies, we acknowledge that controlled manipulation of the trust scores is still possible. In the future, we will combine secure trust logs and lightweight blockchain-based attestation to provide stronger tamper resistance.

Dataset explanation

1. NAB

The Intrusion Detection System receives training and evaluation through the NAB dataset for its operations. Real-world streaming sensor outputs form the basis of time-series data that serves as the dataset type. The platform uses data simulation to replicate edge devices and finds Denial of Service (DoS) attacks and other anomalies. The dataset offers various current–time anomaly signatures suitable for developing minimal edge-based models.

• 2.UNSW-NB15

The system fulfills its function by targeting artificial network intrusions for multi-device intrusion detection and prevention. The attack records are categorized as Fuzzers, Exploits, and DoS, among others. The dataset extends to NAB, as it provides additional security threats that enhance the training model’s generalization.

• 3.Usage in AI-SET

The Intrusion Detection System based at the edge relies on these features for training and validation. The time-based and protocol-focused overview enables users to find anomalies effectively. Table 2 shows the various metrics involved in AI-SET.

Table 2.

Detection metrics.

Metric	Value (%)
Accuracy	94.3
Precision	92.5
Recall	91.0
F1-score	91.7

The research paper¹⁴ summarizes the main performance indicators of AI-SET. The framework’s real-time edge detection and intrusion classification yield high detection scores.

As shown in Table 3, the model accuracy rate grows through successive communication periods in a federated learning system. The system demonstrates its ability to integrate disparate data points during operation through its AI-SET federated learning functionality.

Table 3.

Federated learning rounds versus accuracy.

Rounds	Accuracy (%)
5	65
10	72
15	78
20	83
25	86
30	88
35	89
40	90
45	91
50	91.5

According to Table 4, trust-based access control shows its performance outcomes with varying trust thresholds. The prevention of unauthorized access becomes more effective with higher trust threshold values.

Table 4.

Access control results.

Trust score threshold	Unauthorized attempts blocked (%)
0.2	45
0.4	62
0.6	74
0.8	87

Table 5 compares bandwidth usage among the three configurations. While operating with encrypted data, AI-SET’s bandwidth consumption is minimized at higher levels than cloud-based systems.

Table 5.

Bandwidth usage comparison.

Method	Bandwidth used (MB)
Cloud-based	800
AI-SET (Unencrypted)	560
AI-SET (Encrypted)	630

According to Table 6, trust evaluation operates with maximum efficiency among the three components, enabling instantaneous access decisions. Table 7 shows the comparison between the proposed system and the existing system.

Table 6.

Latency (ms).

Component	Average latency
IDoS	150
FL update	300
Trust evaluation	90

Table 7.

Comparative analysis of AI-SET versus Existing methods.

Feature / metric	Cloud-based IDS	Centralized FL	Edge IDS	FL for privacy preservation	Context-aware access control in fog/edge	Blockchain with edge AI	AI-SET (Proposed)
Anomaly detection accuracy (%)	88.4	89.7	90.8	90.2	89.5	91.1	94.3
Privacy preservation	✖	Partial	✖	local FL	Partial (context-aware)	blockchain-based	with DP + HE
Model convergence (Rounds)	50 +	40	35	35	N/A	33	28
Unauthorized access blocked (%)	N/A	N/A	68	N/A	72	78	87
Latency (ms)	> 500	> 400	200	190	230	160	< 150
Bandwidth usage (MB)	800	700	640	590	620	580	560
Trust-based access control	✖	✖	✖	✖	Static policy	✖	Dynamic + ML

Nevertheless, compared to the classical variants of the IDS, the recently introduced Edge-Resident Intrusion Detection System features several innovations that are specific to the IoT edge environment. To begin with, it employs a hybrid detection system that combines rule-based detection alerts with adaptive AI forecasts, enabling context-sensitive responses to threats. Second, a model-adaptive resource controller is introduced that allows for reducing the LSTM depth and adaptively utilizing edge-node resources (CPU and memory) available in a decision tree. Third, truncated backpropagation and quantized weight layers minimize the energy consumed and ensure temporal detection fidelity because the LSTM model is optimized. A multi-modal set of features that the system is also endowed with includes time-series metadata, as well as machine-specific behavioral indicators that are sensitive to stealthy or zero-day attacks that manage to slip through an ordinary rule-based filter. They are minute advancements that overcome the computing and environmental constraints imposed in the real world, particularly in terms of edge computing implementation on IoT systems.

The AI-SET framework receives a side-by-side analysis against three architectural schemes reported in the literature: a traditional cloud-based intrusion detection center, federated learning, and a state-of-the-art edge IDS¹⁵. AI-SET substantially enhances anomaly detection precision by integrating distributed intelligence capabilities with encrypted data learning methods across different networking points. Through its targeted approach, AI-SET offers improved system adaptability, rapid learning speed, and reduced latency, while ensuring bandwidth protection through state-of-the-art cryptography methods. The research confirms that AI-SET provides an all-inclusive intelligent security framework for IoT systems that can operate efficiently in distributed practical networks.

Figure 3 shows how the federated model learns across 50 rounds. according to the curve stability, the algorithm achieves efficient convergence after 30 rounds.

Fig. 3.

Federated learning accuracy over rounds.

When trust thresholds are adjusted, the system blocks unauthorized attempts, as shown in Fig. 4. Higher thresholds are more secure.

Fig. 4.

Access control effectiveness at various trust score thresholds.

The bandwidth efficiency of AI-SET’s edge-based processing becomes clear in Fig. 5 compared to traditional cloud-based architectures.

Fig. 5.

Bandwidth usage comparison across methods.

Each core component in Fig. 6 presents its average processing delay information. The trust evaluation process occurs the fastest, followed by IDoS and FL evaluation.

Fig. 6.

Average latency of AI-SET components.

The relationship between precision and recall in Fig. 7 demonstrates a high level of model detection accuracy and sensitivity.

Fig. 7.

Precision vesus Recall for AI-SET detection performance.

The evaluation in Fig. 8 demonstrates how AI-SET performs in comparison to current IoT security frameworks, using metrics that measure accuracy, unauthorized access blocking, latency, and bandwidth usage. AI-SET outperforms other solutions regarding all four measurement parameters^16,17. The solution offers superior detection capabilities, resulting in enhanced precision, maximum unauthorized access blocks, reduced latency, and lower data transmission costs due to its edge-level functionality. AI-SET proves to be an advanced and adaptable Internet of Things security solution through this display of performance data. The visual comparison using bar charts demonstrates how AI-SET performs in comparison to current methods in terms of accuracy, latency, access control, and bandwidth utilization. Most security aspect categories demonstrate the superiority of AI-SET, confirming its beneficial edge-enabled IoT security functionality.

Fig. 8.

Comparative performance chart.

Discussion

Evaluation results proved that the AI-SET framework surpasses traditional IoT security solutions in every operational aspect. Security at the edge is built by merging anomaly detection with federated learning protocols that safeguard privacy and trust-based access control systems. This combination delivers full edge-native security protection. Together, the individual modules function effectively on their own and generate combined advantages that enhance the performance level of the entire system. The deployed anomaly detection engine running on limited edge devices delivers 94.3% detection accuracy and minimal false positive results. Model weights are reduced to a minimum, creating effective solutions for edge environments, which enables the deployment of this system across various hardware systems¹⁸. The federated learning component successfully converges within 28 rounds, significantly faster than traditional models. The addition of differential privacy and homomorphic encryption does not affect performance, affirming that robust privacy guarantees can coexist with computational efficiency in real-world IoT deployments.

Although we did not conduct a power consumption comparison as a primary concern, we are aware that power consumption is a significant issue for edge deployment. We will measure the power consumption of various AI-SET component subsets in the future using such tools as PowerTop and Jetson Stats. We also aim to implement adaptive offloading strategies that can dynamically enable energy-aware execution of models on the device, by the device’s capabilities and the workload detected at any particular time, thereby enhancing operational sustainability in low-power IoT-based systems. We had assessed the algorithmic complexity of the main AI-SET components. The anomaly detection module (LSTM-based) requires a time complexity of O(n·t·h) and a space complexity of O(h²), where n is the size of the inputs, t is the number of time steps, and h is the number of hidden units. The round-time complexity of federated learning is O(k.d), where k is the number of devices and d is the dimensionality of model updates. The computation of the trust score is O(m), where m is the number of features of context and behavior. Such low-to-moderate scalabilities are useful and apply to narrow resource scalability on edge servers.

The trust-based access control mechanism represents a significant advancement, dynamically adjusting access permissions based on device behavior and contextual analysis¹⁹. This reduces the risk of insider threats and offers granularity and responsiveness that static ACLs cannot achieve. Blocking 87% of unauthorized access attempts underscores the capability of AI-driven trust metrics in securing complex environments. Comparative analysis confirms that AI-SET outperforms legacy systems in key metrics, including latency, bandwidth usage, and privacy enforcement. It proves especially well-suited for environments where rapid, decentralized decision-making and data confidentiality are critical. The AI-SET framework presents a compelling solution to the multifaceted security challenges encountered in IoT ecosystems, serving as a benchmark for future edge-integrated security architectures. AI-SET I’s scalable, utilizing lightweight models and modular deployment on edge nodes to support a growing number of devices. We evaluated the system’s performance with a maximum of 100 simulated IoT devices, achieving stable trust evaluation and model convergence. Such an evaluation in practice will be the subject of follow-up work to assess long-term scalability in heterogeneous workloads in real-life, large-scale deployments, such as in smart cities and industrial IoT applications. Practically, the AI-SET components must be highly coordinated to achieve compatibility and real-time use. To separate and control the individual subsystems, which include IDS, FL, and access control, we adopted a modular and container-based component deployment approach (utilizing Docker) that maintained data flow among components by utilizing common message queues. Real-time inference synchronization was achieved through lightweight middleware services, and communication bottlenecks were mitigated by utilizing binary data formats and local processing at the edges. This architecture is capable of providing seamless hardware-software interaction and additionally establishes the basis for its deployment in heterogeneous IoT environments.

Limitations

Although its effectiveness has been proven, the AI-SET framework has some limitations. To begin with, although it works great on resource-constrained platforms (e.g., Raspberry Pi, Jetson Nano), ultra-constrained IoT nano is not compatible with all of its features without additional optimizations. Second, applying differential privacy and homomorphic encryption introduces computational overheads, which can influence the performance of latency-sensitive implementations. Third, although the trust model can adjust itself well to observed behavior, the possibility of using the model effectively in cases of adversarial manipulation or large non-IID data settings remains a research question. The above drawbacks present exciting directions for exploration in future efforts aimed at lightweight model optimization, adversarial resiliency, and self-trust calibration.

Conclusion

The AI-SET framework represents a significant step forward in addressing the multi-dimensional challenges of IoT security, particularly in edge computing contexts. By combining lightweight anomaly detection, privacy-enhanced federated learning, and adaptive trust-based access control, AI-SET delivers a robust, scalable, and intelligent solution tailored for distributed and resource-constrained environments. Experimental results validate AI-SET’s superiority over conventional cloud-based or static models in terms of detection accuracy, latency, bandwidth efficiency, and privacy preservation. Federated learning ensures that private data remains local to devices, while homomorphic encryption and differential privacy further secure model updates during collaborative training. AI-SET’s trust evaluation and dynamic access mechanism significantly reduce unauthorized access and improve system resilience against evolving cyber threats. Furthermore, its modular and scalable design ensures compatibility with various IoT platforms and devices.

Future work

Building on the promising results of this study, future research will explore:

• Integration of reinforcement learning for continuous self-adaptation of access policies

• Real-time threat intelligence sharing between edge clusters

• Deployment in industrial-scale environments and smart city infrastructures

• Hardware-level optimization and deployment on ultra-constrained devices (e.g., MCUs)

• Formal security verification of federated models under adversarial settings

These directions aim to evolve AI-SET into a fully autonomous, self-learning edge security agent capable of proactive defense in future cyber-physical systems.

Although AI-SET takes the form of a general-purpose security framework, we acknowledge that certain domain-specific adjustments are necessary. The next steps will involve tailoring the system to address the concerns of critical industries, including the healthcare industry, where compliance with regulations such as HIPAA and fine-grained data access auditing is mandatory, and the industrial IoT, where the ability to respond to threats within low-latency or high system resiliency is essential. It will include adjusting trust models, access policies, and training routines to conform to the operational limits and specific threats of a given sector.

Author contributions

V. Padmavathi: Problem Selection, Algorithm, implementation, Results, coding, testing R. Saminathan: Experimental Results, Algorithm part and Editing.

Data availability

The datasets used in this study are publicly available. The Numenta Anomaly Benchmark dataset can be accessed at https://github.com/numenta/NAB, and the UNSW-NB15 dataset is available at https://research.unsw.edu.au/projects/unsw-nb15-dataset. For further information or requests regarding data usage in this study, please contact the corresponding author, Dr. V. Padmavathi, at vvpadhuavc@gmail.com.

Declarations

Competing interests

The authors declare no competing interests.