A lattice-integrated AES framework for ultra-secure biometric protection on resource-constrained edge devices

Abstract

In recent technological advancements, consumer electronics such as mobile phones and edge devices have expanded rapidly. Still, critical security issues related to the confidentiality and integrity of user data is persistent. Over 1.2 billion consumer devices are in use, with varying user data. Still the secrecy of user biometric information, such as fingerprints, voice, and facial mapping is being compromised, leading to significant losses for consumers and their valuable savings. To overcome this, the study introduces a fingerprint-template protection framework to keep user biometric information safe on consumer electronic devices, preventing tampering with user data. The proposed methodology is lightweight and runs smoothly on resource-constrained devices, with an effective feature extraction and normalization of user fingerprint data. The model utilizes AES-256 Galois counter mode (GCM) with PIN oriented keys and a lightweight lattice-based randomness generator to secure user information. The proposed framework achieves a fake fingerprint detection accuracy of 98.69% with minimum turnaround for encryption and decryption, compared with other state-of-the-art methods such as RC5 (59.6%), ElGamal (91.1%), and RSA (92.35%). The experimental results show that the proposed framework is superior in both security and data accessibility. Integration of lightweight lattice-based randomness generators in consumer electronics improves security and enables more reliable on-device applications.

Introduction

Consumer behaviour data refers to information collected from customers’ actions, patterns, and preferences during interactions with products and services¹. This data includes Demographic information, consumer transactional behaviour, online activity information, social media data, location information, and customer service interaction information. This information is beneficial for personalisation, enhancing customer behaviour, improving product development, optimising marketing strategies and predicting customer trends. Protecting this consumer data is very important for many reasons, including privacy concerns, legal commitments, and the risk of data breaches². They are not only protecting sensitive information but also maintaining customer trust and mitigating high risk. Multi-Factor authentication is one of the most effective mechanisms for protecting this data, requiring more than one form of verification before granting access to consumer data³. Likewise, Zero Trust methods, Secure data deletion, Hashing, and anomaly detection methods can be used for these data protection measures⁴.

Biometric authentication technologies, primarily fingerprint recognition, use unique psychological characteristics to identify and verify individuals⁵. Incorporating fingerprints has become increasingly essential for consumer data protection due to their strong security. Fingerprint authentication is used for access control to sensitive systems and data⁶. Also, this method is a user-friendly and highly secure way to protect customer data from unauthorised access. To reduce the risks associated with traditional authentication methods, such as password attacks, fingerprint templates are increasingly being used for security functions that cannot be duplicated⁷. Even though fingerprints can provide strong protection, there are many challenges, such as fingerprint template leaks. Sometimes, fingerprints can misidentify users, leading to access control issues. If the system used high-quality fingerprint sensors, then problems can be minimised⁸. Hardware dependence is another challenge for fingerprints. Because not all devices may have fingerprint sensors, it may be inaccessible. Privacy concerns are another challenge with fingerprints. As fingerprints are unique and tied to individuals, they are challenging to store securely and to comply with protection regulations⁹.

The main problem with the biometric authentication framework is its irreversible, non-reversible, and permanent nature, which raises concerns and makes templates easy for an imposter to imitate. This template protection technique is based on fingerprint feature transformation and key binding, with limiting entropy fingerprint sources, which are vulnerable to offline attacks and privacy issues due to side-channel leakage. Devices that utilize user PINs and key derivation are prone to brute-force, replay, and leakage attacks on resource-constrained smart devices. The primary motivation of this research is to consider the above problems by processing a secure fingerprint template with a lattice-based randomness generator for key generation by integrating the user PIN with learning with error (LWE) entropy and KDF2 key for safe and heavy cryptographic functions for achieving fingerprint matching accuracy compared with the original and the imposter. For high-performance analysis, the proposed work utilizes the FVC2002 dataset and conducts an ablation study compared with other databases.

The proposed methodology is planned to be developed through a structured multi-phase approach that combines biometric data processing and fingerprint authentication systems using datasets such as the FVC2002 databases, with a single fingerprint dataset. This will be optimised for consumer devices. Biometric traits, such as fingerprints, will be collected during the data collection phase. The proposed methodology used lightweight encryption and a random number generator to protect fingerprint templates. Here, AES-256 Galois Counter Mode (GCM) is used as a lightweight, similarity-based fingerprint template authentication method rather than real fingerprint image classification, as a ciphertext modulation. The methodology of the proposed work includes a plain minutiae fingerprint template for reconstructing biometric data on consumer devices. The template and AES process efficiently prevent user data leakage on resource-constrained devices without downgrading fingerprint detection accuracy. The limitation may arise because symmetric encryption does not prevent replay attacks, such as those involving user-specific, guessable keys on smart devices. To overcome this limitation, the proposed methodology utilizes PIN-based key generation with a template to verify user biometric data. For additional security, a lattice-based random generator is used in the AES-GCM architecture. This random generator is not a heavy post-quantum lattice system; it produces a standalone cryptographic random generator, making it lightweight, deployable, and free of additional computational overhead.

The evaluation of the proposed methodology mainly focuses on accuracy and computational time. If accuracy is higher and computational time is lower, real-time operation with minimal memory usage is possible. This framework will balance accuracy and time. This will perform well in standard template protection systems, where both original and spoofed fingerprint scores are generated via cross-fingerprint identity pairing. This process increases verification across original fingerprint variability analysis, rotation, and partial fingerprint pressing on smart devices. The evaluation parameters are within the threshold limits, and the resources are authenticated in real time.

As consumer behaviour data is crucial to many business activities, protecting it is mandatory to avoid legal and ethical issues. There is a need for a fingerprint-based data protection system that prioritises security, usability, and privacy.

Contributions

A biometric authentication system that preserves the integrity and privacy of the consumer device data has been proposed. The method combines the features of cryptography and computing. The contributions of this study are as follows:

1. Developing highly secure fingerprint template protection that integrates fingerprint with lattice-based randomness generator for efficient cryptographic user key generation.

2. Combining LWE based salt with KDF2 and AES-256 GCM to improve entropy to product offline attacks and side channel linkage issues.

3. Development of a strengthened secure key generation mechanism using SHA512 with lattice-based randomness generator for securing fingerprint data.

4. Processing efficient experimental analysis using FVC2002 dataset to attain high security template protection.

5. Demonstration of the feasibility of integrating advanced cryptographic techniques for securing user data against different attacks in consumer electronics.

Organization of the paper

The remaining part is organized as follows. The section II explains the related works and the section III presents the material, methods and working procedure fingerprint feature transformation for the key generation model. The section IV explains the proposed work and the section V presents experimental setup, experimental evaluation and performance analysis. The Conclusion and future work have been discussed in section VI.

Related works

The related works focus primarily on the key areas like AI based security models, sensor-based models etc. to ensure the user privacy protection in the consumer electronics.

AI based security models

Tran et al.,¹⁰ has discovered a new method which includes AI based classification along with normal enrolment and authentication. They argued that the existing schemes are vulnerable to the biometric authentication which is deployed in mobile devices. Their proposed methods contain an authentication method which includes an AI based binary classifier. This classifier will classify the biometric data which includes Iris and fingerprints. They evaluated the proposed method with FVC2002-DB1, FVC2002-DB2, FVC2002-DB3, FVC2004-DB2 and UBIRISv1 datasets and reduced the Equal error rate 21%, 23%, 17%, 24% and 26% respectively.

The research presented by Francesco Antognazza et al.¹¹ used the post-quantum security function with the use of the lattice cryptosystem. Without the lattice cryptosystem, the hardware used for the particular system is prone to confidentiality attacks like man-in-the-middle; this leads to a huge number of attacks being processed towards the system. To overcome these attacks, lattice-based cryptography is used to provide security and efficiency towards hardware systems to perform computation efficiently to achieve confidentiality. The model uses AES-128 as the security level for the design of X-Net for efficient learning of the system. The suggested work has taken more time to compute a security operation, as it is more suitable for quantum-based devices. Sanghoon Lee and Ik Rae Jeong¹² have proposed a new algorithm for fingerprint matching algorithms with less time. According to the authors, the size of the individual fingerprint is smaller than the fingerprint database. So, they have proposed an indexing method for the fingerprints which will take less time in the databases. The Hit ratio of the proposed method is 92%. They have suggested that the proposed algorithm can be integrated with machine learning approaches for the widely spread distributions.

Stefano Marrone and Carlo Sansone¹³ proposed a method for perturbation attacks using CNN. The perturbation attacks will introduce small changes in the input data which leads to misclassification or incorrect predictions. The authors have proposed an authentication system which takes fingerprint data as an input and verifies the attacks. The convolutional neural networks will be used for the detection of liveliness of the fingerprint data. Here, the attacker will use major approaches namely image resize and noise resize for changing the image sizes. The proposed method contains three major steps namely preprocessing, registration and minutiae pairing. The crossmatch images using Deep Fool tool has given 87.01% efficiency and FGSM tool has given 88.44% efficiency.

To enhance the authentication further, Sumalatha et al.,¹⁴ have identified a homomorphic deep learning-based framework using fingerprint authentication. In the proposed approach, the fingerprint authentication system will integrate the deep learning and homomorphic encryption method. The authors chose the dataset from SOCOFing which contains 6000 fingerprints from 600 individuals. After the encryption technique, the CNN will be applied and the authors have calculated the accuracy. The proposed method gave 97.9% accuracy and 81% false rejection ratio which is computationally feasible when compared to existing methods.

Sensor and IoT based security models

Jae Yeol Jeong et al.¹⁵ extended the Roy et al., work for the small fingerprint sensors. They analysed that the size of the partial image is 15% then it can be possible to produce a Master Print of the fingerprints. They have proposed a Sample Master Print Generation algorithm for this process and they have achieved 91% which is comparatively good. Deepika Gautam et al.¹⁶ presented a biometric based authentication system for Internet of Drones. To overcome the issues in the existing methods, the authors have proposed a novel biometric based authentication system which integrates physiological behaviours to strengthen identity verification and enhance resistance to common attacks. The performance evaluation is done to compare the computational and communication overhead.

Bushra Khalid et al.¹⁷ presented solution for key agreement scheme using biometric data for intelligent sensor based wireless communication. The authors have presented a model called Biometric based User authentication and key agreement protocol which uses a one-way hash function for authentication and fuzzy systems. This model uses the fuzzy extractor to update the biometric template. It uses generate and reproduce modes for employing the datasets. The major steps of this method are setup and registration. The computational cost of the proposed model is very less when compared to existing models like AKA and sensor based SLUA.

The sensor-based authentication in the smartphones is an unconventional approach which controls the smartphones to verify the identity. Unlike the traditional methods like passwords and PIN, the sensor-based authentication enhances security and personalization. Moceheb Lazam Shuwandy et al.¹⁸ suggested a model for this sensor-based authentication system. The authors have discussed the various types of sensors like fingerprint sensor and touchscreen sensor along with their operations. Also, they have discussed the security impact on smartphones, sensor authentication usability and sensor security along with the challenges.

Manasha Saqib et al.¹⁹ proposed a lightweight three factor authentication framework for IoT based applications. In the proposed model, the authors gave much importance for the signature generation and verification which is based on the biometric data. They have strongly believed that the proposed methodology is very secure against the attacks and it will ensure confidentiality, authentication and data integrity. They have estimated the computational cost of the proposed methodology is 0.190 s and the computational overhead is 2560 bits.

Ole Höfener and Qinghua Wang et al.²⁰ analysed the requirements to build a framework for device authentication using fingerprint data. The authors have considered many parameters like feature availability, feature stability, feature uniqueness, accuracy etc. which contribute more in the framework construction. The constructed framework can be used from high-level to low-level IoT devices. Also, it will give more sophistication against the attacks.

Computing based security models

Quang Nhat Tran et al.²¹ discussed various privacy preserving mechanisms which includes non-invertible mechanisms, Information hiding techniques, Protocol based protection and Direct biometric key generation. They have taken the dataset which includes fingerprints, face and iris. They have used the secure multiparty computation-based biometric security protocol and they have achieved 94% accuracy rate.

Mohammed Hammad et al.²² presented a cancellable multibiometric model which can be used for fingerprints based psychological and behavioural traits. For this study they have used NUPT-FPV dataset which includes fingerprint authentication and finger veins. They have identified the salient features from the dataset and constructed a non-negative matrix factorization model. This consists of a data matrix, a basis matrix and an encoding matrix. The main idea of this model is to create a coefficient matrix so that we can generate a cancellable model which will provide biometric authentication. The authors have achieved 90.5% accuracy in the fingerprints data and 70% accuracy in the finger vein data. Baghel et al.²³ proposed a biometric based authentication system which is mainly used for consumer electronic devices. They have used a Discrete Fourier Transform with a 3D grid which will be useful for diversity and security. They have taken the FVC2002 DB1 dataset and estimated the equal error rate which is 86%. This showed the superiority of the results and effectiveness.

Authentication based security models

Sumalatha U, et al²⁴ reviewed the advantages and disadvantages of fingerprint recognition systems after the performance analysis. They have taken the advancements of biometric technology for this study. They have proposed a fingerprint authentication model which includes enrolment task, verification task and identification task. With this model they have estimated the equal error rate as 11% which is significantly low when compared to the existing methods. The research coined by ZEYAD GHALEB AL-MEKHLAFI et al.²⁵ utilized lattice-based cryptography to authenticate vehicles by using the 5G network. The lattice-based encryption utilizes the trusted authority function by transforming multiplication and provides security over quantum-based attacks. The proposed research utilizes AES to transfer the anonymous data towards base station communication. The above research work has the advantage of minimum computational time for single-user verification at 0.3149 ms, but this work has higher transmission costs and is not suitable for consumer-based, resource-constrained devices.

The work done by Zexiang Zhang et al.²⁶ applied an advanced lattice-based Fourier transform for resource-constrained devices. The initial encryption process uses AES to reduce the size of the key ‘s ciphertext. To fasten polynomial multiplication, the Fourier transform is used additionally to increase precision. The work demonstrates less computational time and less memory consumption for the Falcon-based implementation model. The above work has the disadvantage of consuming more memory during the signing phase, which increases the memory overhead in the multiplication function and takes more time to authenticate the user in consumer electronics. Emanuela Marasco et al.²⁷ investigated the biometric multi factor authentication using Finger PIN scheme. The authors have proposed a model with three major steps namely enrolment, registration and authentication in challenging mode. They have used additional and different components for the above steps as it is a multifactor authentication. The dataset was collected from George Mason University which contains 100 subjects from Northern Virginia university. They have achieved the efficiency of the multifactor authentication system of 89.78%.

U.R. Saxena and T. Alam²⁸ reviewed RSA as a homomorphic encryption method, which is used for cloud-oriented fingerprint storage, where RSA provides secure key management with high protection functionality. Still, RSA is computationally expensive, and its encryption time for a smaller key size is also high, exceeding 20 s, making it less suitable for real-time consumer electronic devices. H. Abroshan²⁹ proposed ECC & Blowfish, which provide a shorter key size with a strong security function. However, implementing this fused framework requires additional development time to achieve the desired result and high energy consumption to generate the final result, making it infeasible for low-resource-constrained devices.

The work done by H. Huang and H. Zong³⁰ used hormonal encryption for a biometric matching scheme based on both fingerprint and iris domains. However, during the cryptanalysis process, latency increases above 1 s for each biometric match, which limits its application implementation on consumer electronics. S.H. Murad and K.H. Rahouma³¹ demonstrated the advantages of AES in implementation on resource-constrained devices due to its low energy consumption. They are mostly applied in various fingerprint security applications, offering low computational overhead when combined with other models, such as SHA keys, for improved latency and suitability for real-time implementation in consumer electronics.

Template based security model

The research by Risto Vaarandi and Hayretdin Bah³² highlighted the advancements in template data security, highlighting that many traditional models utilize large language models, which are primarily unsupervised in nature. For handling unsupervised LLM, some specialized templates are used to detect cyberattacks, which many focus on device security logs, not on biometric or user-based data. The advantage of the defined work is that it handles security logs effectively by achieving a reasonable F1 score of 78.3%.

In their research, Ali Hameed Yassir Mohammed et al.³³ proposed the security of biometric data in templates; the iris template is protected with the help of a ranking system. The ranking system employs a hybrid approach that combines DenseR and OrdinalR as a template protection function. The defined hybrid model has a complexity of O (n log n) and a data loss of 0.4402%, which ensures that the system is authenticated using the iris protection template, achieving a detection accuracy of 0.92%. The ranking system securely stores iris data and retrieves it when needed. The research coined by MOHD IMRAN et al.³⁴ proposed a random projected cancellable biometric template, which stores all the user-related information. The information is stored in a hash template of a matrix where data is transferred into different biometric records. The model uses a different fingerprint dataset, which maps six different records to minutiae information about all users. The model achieves an accuracy of 84.2% in predicting a user’s fingerprint using the MCC key; however, the mentioned work is vulnerable to adversarial attacks.

Technical gap

The technical gaps identified from the above literature are expressed below.

1. Modern security schemes reduce biometric data matching accuracy with template protection schemes.

2. Traditional approaches such as ElGamal, RSA, and RC5 are not fully compatible with resource-constrained consumer electronics like smartphones. This led to the lowering performance of the models in low-memory devices with mismatched detection on real-time application security.

3. The traditional model has few systems integration of pin-oriented authentication with randomness generators.

4. The final technical gap with the traditional security model is the lack of multi-layered security to secure spoofing attacks and hashing protection fail due to minutiae variability on user biometric data.

Materials and methods

This section illustrates the datasets, algorithms used and procedures for developing and evaluating the proposed security framework for protecting the biometric data on the consumer devices. This also verifies and validates the efficiency and deployment ability of the proposed framework across various consumer devices to ensure the reliable user data protection.

Materials

The dataset has been collected from databases namely FVC2002 DB1. It contains fingerprint verification Competition 2000 samples with 22 author’s own fingerprints where 100 subjects with 8 impressions per subject. This data primarily focuses on fingerprint data in grayscale images with a pixel size of 388 × 374 utilized optical sensors with 500 dpi. These pixel images provided for feature extraction with raw data. These biometric data can be used to represent in the form of user identification like fingerprints. This data is additionally simulated using a random generation method. The dimension of the data is a 160 × 160 matrix which represents the random distribution of Biometric features (between 0 and 1). This dataset will mimic the real biometric data and it is very useful to test the cryptographic models. The sensitive data in this model can be used for encryption and decryption. It represents personal or confidential user information like secret messages. All performance evaluation function is taken from the FVC2002 dataset it is selected for high quality fingerprint impression without any consideration factor to alter modification by artificial implications. The preprocessing function can be carried out towards FVC 2002 with greyscale normalization and feature extraction to ensure constant validation across each user fingerprints samples. This experimental protocol ensures no user fingerprint overlap on enrolment and validation which secure performance evaluation bias. The encryption and decryption of this secret data will be performed by AES method where the key is generated from the biometric features and user provided PIN suitable for minutiae-based verification³⁵.

Methods

The Biometric fingerprint Data Generation method has been proposed for this study which is used to simulate biometric datasets. These datasets will be processed and validated using the cryptographic operations. The feature extraction will be processed with the fingerprint images³⁶. The images are separated as 16 blocks to compute variance for removing the noise. In addition to this normalization process is handled to ensure clarity of fingerprint image after enhancement expressed in Eq. 1.

1

where M, V as mean and variance, M_0, V₀ as normalized constant³⁷.

A Gabor filter is used to identify the different structure in fingerprint image this will enhance the minutiae-based fingerprint template stability for improving fingerprint matching accuracy expressed in Eq. 2

2

where x, y are coordinates, as wavelength biometric data.

After fingerprint image enhancement adaptive otsu thresholding that converts images into values as 0 and 1then minutiae extraction is processed with cross number methodology expressed in Eq. 3.

3

where P_i is referred as binary neighbours where noise minutiae is removed with distance filtering of each binary numbers.

The fingerprint template format which is complaint ISO standard where each minutia encodes and processes expressed in Eq. 4.

4

where n is number of minutiae template fixed 512-byte structure for achieving compact storage, quick encryption with AES blocks.

After the block processing, the next process is function learning with error (LWE) lattice as a lightweight methodology not performing encryption but to generate randomness for AES-GCM, KDF salt for session oriented random values expressed in Eq. 5.

5

where A is pseudo random matrix, s as secret vector, e as gaussian error and q as modulus of 4096.

The above function provides high entropy which assures unique nonces and collision avoidance salts effectively suitable for resource constrained devices which secures the side channel leakage. LWE has high dimensional lattice which only give limited information this leads to worst case lattice generation problem like GapSVP more protective against side channel leakage attacks and secure user template information.

In the proposed methodology, the PIN is not used in isolation for key generation. Instead, it is hashed via KDF-SHA-512 for one time authentication which is not stored inside the smart devices expressed in Eq. 6

6

where slat derived for learning with error-based noise,100,000 iteration for protection over brute force attack, 256 bits AES key as total output length.

After PIN derived function is processed the AES-256-GCM encryption fixed for each fingerprint template. Initially a nonce with LWE randomness generator is used to create AES key for each PIN then it applies to 512-byte template to store cyphertext and authentication tag expressed in Eq. 7.

7

Each fingerprint template storage size is below 1 KB.

During final authentication user enters the fingerprint where AES-GCM decrypts the fingerprint template for extracting the minutiae which compares live capture fingerprint template for final matching expressed in Eq. 8

8

The biometric features are first passed through a Lightweight lattice with PIN for robust encoding. The output is combined with the KDF2-SHA-512 hash of the PIN, ensuring non-reversible, tightly bound multimodal features which are highly specific to the user and session. The major advantages of this method include enhanced security, Improved performance, privacy reservation, scalable framework and practical adaptability. Overall, this study is a futuristic model to secure the fingerprint data in the consumer devices^38,39.

Proposed methodologies

The proposed model aims to preserve the integrity and privacy of consumer devices using lightweight learning with error (LWE) lattice randomness generator with AES-GCM method which takes key value from the biometric feature and user’s PIN.

A Fig. 1 Explains the overall process of the proposed model Flow. The proposed work uses the lightweight learning with error (LWE) and lattice randomness generator with AES-GCM based advanced encryption technique which combines the biometric features and user input. This will create the cryptographic key as stronger⁴⁰. For biometric data processing, the user’s biometric data is like fingerprints. This will be processed using a biometric feature extractor. This will create a digital representation of the biometric data. To generate a strong cryptographic key, a hash function will be applied on the digital data with a key length of 256 bits. Meanwhile, the user input PIN will be stored separately as a SALT value. Using concatenation operation, the biometric data and the user PIN will be combined. This combined input value will help for the key derivative function in AES-GCM⁴¹. Once the key value is generated with KDF-SHA512, the biometric data will be encrypted with AES strong mode. For the decryption, the user should provide the scanned encrypted biometric data along with the PIN value. The verification authority (VA) ensures that the data differentiation between the fingerprint PIN generated by the user’s consumer devices and the decrypted template retrieved from secure data storage is maintained. This enables the verification authority to ensure data integrity before allowing user access, with clarification of incoming data from consumer devices or storage databases, thereby improving classification accuracy⁴².

Fig. 1.

Overall schematic flow of proposed fingerprint template by integrating lattice randomness generator, Key derivation and AES GCM encryption.

The dataset used in the proposed model contains the author generated synthetic fingerprint images which is created using a controlled pattern generator which models ridge orientation. A subset of real fingerprint images sourced from the FVC2002 database, which served as growth truth and benchmarking references. All images were normalized to a pixel size which is synthesized at 500 DPI resolution size.

While synthetic fingerprints were used primarily to test the noise resilience and feature transformation robustness of the lightweight learning with error (LWE) lattice randomness generator with AES-GCM, the generation will be guided by visual validation through overlays with sample real prints from FVC2002. These results show that the synthetic images are structurally comparable to real-world patterns especially in capturing ridge valley consistency and minutiae dispersion⁴³.

Working procedure for the proposed method

The following section, it has described that the steps involved in the proposed method.

Step 1. Fingerprint data acquisition as extracts minutiae and do preprocessing with 500 DPI with pixel intensities in 0 and 1 encodes them into fingerprint template towards consumer electronic device.

Step 2.After data acquisition, normalization of biometric fingerprint to size of 388 × 374 pixel from FVC dataset with Gabor enhancement increases the clarity and reduces the image noise.

Step 3. After normalization of synthetic biometric fingerprints, the next input is to do minutiae, template using cross number classification which cleans the fingerprint co-ordinate and image orientation.

Step 4. Process minutiae template into 512-byte ISO template which ensures interoperability and minimizes the computational overhead.

Step 5.LWE with Lattice randomness generator processes random salt with KDF-SHA512 and AES GCM nonce which secure data with cryptography and non-duplicate random values with low computational cost.

Step 6. The user fingerprint transfers into a 256-bit AES key to protect user data in consumer electronics from PIN based brute force attacks.

Step 7. Finally, the template decrypted using AES-GCM any data tampering leads to failure due to GCM tag verification.

Step 8. Minutiae template fingerprint matching is performed with Minutiae cylinder code analyse the fingerprint neighbourhood were genuine matched process device entry.

As the consumer electronics devices keep increasing and integrating with biometric systems for authentication, the proposed model will safeguard the personal data. The reason for using AES-256 GCM is to prevent access of fingerprint data without the correct fingerprint. This makes AES encryption consistent and unbreakable for different biometric fingerprint data. Modification of ciphertext results in GCM authentication, which increases stability against cyberattacks such as brute force and dictionary attacks. Using PIN derived keys before the encrypted fingerprint data keeps it safe; if someone tries to figure out the fingerprint data, they won’t be able to get back to the specific consumer electronics, making the system safe from reverse engineering attacks. Even if two devices store the same fingerprint, the LWE lattice generator merely converts the fingerprint into a hash with different salts which prevents cross matching of fingerprints. Applying KDF-SHA-512 before AES is essential to prevent devices from different attacks with lightweight randomness generator run per each user data authentication effectively in resource constrained devices⁴⁴.

Data generation and preprocessing

The main objective of this step is to generate synthetic biometric data, preprocess it, and extract features for the cryptographic operations⁴⁵. Assume that D_Synthetic is a synthetic biometric data matrix and PIN as user provided for key generation where i and j input data for matrix multiplication, as constant. The Synthetic data will be generated using the following Eq. 9.

9

This uses a uniform distribution to create a matrix representing biometric data as . After generating the synthetic data, preprocessing normalization will be done on the data with minimum and maximum as min, max expressed in Eq. 10.

10

The normalization will be ensuring the values are between 0 and 1. The feature extraction using the AES-GCM can be done with d expressed in Eq. 11.

11

PIN value

To strengthen the framework and provide clarity, the following specifications can be given for the generation of PIN value. The user’s PIN can include numeric digits (0–9), alphabets (A-Z/ a-z), special characters including diacritics (á, é,…) for international support and a total of 94 printable ASCII characters. It can be the length of 6–16 characters which may ensure the usability and security. All the PIN inputs will be normalized using UTF-8 encoding prior to the integration into the lattice matrix and AES256 encryption pipeline. This will help to include the diacritics into the encryption process and maintain the consistency across diverse user locals. The normalized PIN value acts as a dynamic key modifier for biometric encryption which enhances the resistance against brute force attacks⁴⁶.

Cryptographic key generation, encryption and decryption

The goal of this algorithm is to generate a cryptographic key as K based on the user PIN and feature vector and then perform encryption and decryption of sensitive data. To perform this, assume F is the feature vector generated from the biometric data, PIN is user’s information for key generation and S is the sensitive data to be encrypted like user’s personal information⁴⁷. Then the cryptographic key will be generated using Eq. 12

12

This uses the AES-256 encryption algorithm with the cryptographic key K to encrypt the padded sensitive data. The decryption of the encrypted data will be done as follows Eq. 13.

13

This uses the same key K to decrypt the ciphertext C in Eq. 14.

14

This will remove the padding from the decrypted data to obtain the original sensitive data S_unpadded. In Eq. 14

Performance evaluation of cryptographic functions

This algorithm will analyse the performance of the encryption and decryption process, including the execution time and accuracy⁴⁸. For the performance analysis, assume N is the number of test cases, C is the ciphertext generated from the encryption process, S is the Individual sensitive data, and S_unpadded is the decrypted sensitive data. The encryption time T_enc is defined as the time taken for the encryption process in each test case and the same will be calculated using Eq. 15.

15

Like the encryption time, the decryption time T_dec is defined as the time taken for the decryption process in each test case in Eq. 16.

16

The average encryption time and decryption which takes for encryption and decryption across all times can be defined as Eq. 17 and 18.

17

18

After the calculation of average encryption time and average decryption time, the accuracy should be calculated. This will compare the original data S with decrypted data S_Padded to verify whether the encryption and decryption process is accurate with verify authority VA^49,50.

19

where TP represents the predicted values are correctly predicted as actual matches. TN represents predicted values that are correctly predicted as an actual mismatch. FP means predicted values are incorrectly predicted as matched. FN means positive values correctly predicted as actual mismatched is expressed in Eq. 19.

Final integration

Once these algorithms are implemented, then.

• • Algorithm 1 generates the Fingerprint biometric template data and preprocesses it into a feature vector.

• • Algorithm 2 securely encrypts and decrypts sensitive data based on the generated feature vector and user PIN.

• • Algorithm 3 evaluates the performance of the encryption-decryption process to ensure efficiency and accuracy.

The integration of these algorithms will give the proposed model securely handle biometric data for consumer electronic devices. The system can scale to different devices and networks with performance estimations.

Algorithm 1.

Generate Synthetic biometric fingerprint Data.

Algorithm 2.

Generation of Key.

Algorithm 3.

Encryption and decryption process.

Experimentation, results and analysis

Experimental setup

An experimental setup for the fingerprint-based consumer device data requires the integration of biometric authentication systems into encryptions .The initial setup includes standalone system with processor Intel I5 intel XE graphics card with 16 GB ram, Libraries like PyCharm, TensorFlow and metaplot, environment as Google colab with virtual ram of 12 GB .This includes setting up of sensors, feature extraction, matching algorithm, enrolment of fingerprints, design of encryption and decryption algorithm. The dataset contains 2000 images with image size of 388 × 374. The quality of the image is 500 DPI. The fingerprint dataset which is used in this research contains high resolution grey scale fingerprint images designed particularly for AI-based verification in consumer electronic devices. The dataset is the combination of the fingerprint samples and additional images sourced from the publicly available Fingerprint Verification Competition (FVC) 2002 database. Each fingerprint image has a standard dimension of 388 × 374 pixels and is scanned at a resolution of 500 DPI which aligns the standard fingerprint image resolution used in biometric models. The key features extracted from these images include the texture details, ridge patterns and minutiae points which are essential for reliable fingerprint authentication. The primary objective of using this dataset is to develop and test an enhanced fingerprint verification algorithm optimized for deployment in resource-constrained consumer devices.

Experiment evaluation

The experimental evaluation of the proposed model focuses on accuracy, matching and encryption and decryption time. Also, it is decided that the test environment should be mixed like clean, dry, wet damaged fingers so that the device performance can be estimated under different conditions. Figure 2 illustrates the encryption and decryption flow of proposed model.

Fig. 2.

Proposed model architecture with secure enrolment of fingerprint template encryption and authentication towards decryption.

The Fig. 2, explains the sequences of processes involved in the proposed model encryption and decryption. The consumer uses smart devices like smartphones which will integrate the fingerprint scanner. These fingerprints will be scanned and the features will be extracted. From this, a fingerprint template will be generated. These templates will be stored as an image with a size of 388 × 374. For the key management step, these features will be extracted and concatenated with the user’s pin value. Then the AES-GCM encryption with LWE lattice randomness generator-based fingerprint feature transformation for key generation encryption and decryption will be performed once the key value is generated. This process illustrates fingerprint enrolment, key generation, encryption and decryption. For the performance estimation, the proposed model calculates the accuracy, matched fingerprints, encryption time and decryption time.

This section explains the performance analysis of the proposed model. The Fig. 3, represents the heatmap of the collected dataset FVC 2002 DB1 where 800 images are used for 8 impressions with a single subject. The training and testing are processed with 4 impressions for training and 4 impressions for testing. The genuine pair test is generated with 4 test impressions with total genuine comparison as 4, 2 as 6 for 100 subjects which is 600. Next with imposter evaluation for one impression for all subjects as 100 products of 99 as 9900 this as standard FVC protocol which assures fair testing evaluation.

Fig. 3.

Correlation Heatmap fingerprint template feature vectors over multiple biometric fingerprint samples.

Heatmap

This heatmap is used to visualize the data patterns which represent the relationship between the data expressed in Fig. 3. Correlation heatmap generated with FVC2002 dataset, the fingerprint feature correlation heatmap is plotted using FVC dataset for intra data similarity among fingerprint samples. each fingerprint was pre-processed using histogram equalization. Local binary pattern features are extracted from each fingerprint image for compact lightweight feature vector suitable for consumer electronics. First a pairwise cosine similarity was computed with feature vector resulting a square similarity where each block represent correlation between available fingerprint samples. higher similarity value indicates strong feature correlation similarly low indicate weaker correlation. This matrix is visualized as a heatmap for describing the consistency and fine distinction capability extracted from the fingerprint across the FVC2002 dataset.

Confusion matrix

The confusion matrix is very useful for performance analysis of classification models. For the fingerprint authentication system, the matrix has developed between the predicted and actual classifications. The Fig. 4, shows the confusion matrix generated using FVC 2002 DB1 dataset. Here, 600 total samples as genuine in which 593 correctly identified as genuine, 9900 imposters in which 9776 identified as imposter fingerprint for mismatch attempt. This will help to understand the actual matches and minimize the false matches.

Fig. 4.

Confusion matrix obtained from FVC2002 dataset with differentiation of original and imposter classification on normal operating conditions.

ROC-AUC

The Fig. 5 gives the complete measure of how the system correctly identifies the matches and mis-matches during proposed work internal validation the proposed AES-256 Galois counter mode (GCM) with PIN oriented keys and lightweight lattice-based randomness generator gave 0.99 of AUC towards true positive whereas the false rejection rate is 1.31%. This is comparatively good in the fingerprint security on resource constrained devices.

Fig. 5.

AUC-ROC Curve of the proposed method evaluated on FVC 2002 under internal validation.

Training accuracy

The Fig. 6 plotted between accuracy value vs the samples and it is distinguishing between legitimate and illegitimate users during the authentication verification. This time depends on the algorithm, size of the fingerprints, key value, mode of the encryption and the storage. As the proposed model uses AES-256 Galois counter mode (GCM) with PIN oriented keys and lightweight lattice-based randomness generator, it takes less than 3 ms for encryption with multiple number of iterations. It has proven that the user experience is good with the proposed algorithm.

Fig. 6.

Training accuracy and loss during each epochs defining the trade-off operation with fingerprint security system.

Matched fingerprint complexity analysis

The Fig. 7 illustrates the matched and unmatched fingerprints which are plotted between security levels vs samples. This depends upon uniqueness of the fingerprint data, matching algorithms, encryption algorithm and user enrolment. The computational time to process the proposed work is the time to extract fingerprint features; after extraction, normalize the data, and then the biometric feature extraction takes place at a certain time. After extraction, the next process moves to a lattice randomness generator for fingerprint feature transformation and key generation. After key generation, the KDF-SHA-512 PIN function processed to generate a consistent key where each process complexity as O(n) for minutiae extraction, O (1) for AES-GCM encryption, O (n²) for Lightweight LWE lattice randomness generator, O (N log N) for MCC which reduces the memory overhead. The processing time for encryption is 0.020, and the decryption time for the same regeneration is 0.018 s, which reduces the computational overhead. This minimum regeneration time makes the proposed model suitable for real-time applications as a lightweight process for consumer electronics with no single-stage lattice operation, which reduces the computational bottlenecks.

Fig. 7.

Matched fingerprint security lattice randomness generator for fingerprint feature transformation and fingerprint samples.

The execution time of AES-GCM and Lightweight LWE lattice randomness generator minimizes the memory overhead where minutia extraction as 28.4 ms, ISO template encoding as 4.1 ms, AES 256 GCM encryption/ Decryption as 20.5/18.3 ms and MCC matching as 15.7 ms which is more suitable for consumer electronics authentication with total execution time as below 100 ms on internal validation.

Data encryption and decryption analysis

The Fig. 8 shows the encryption time and decryption time taken based on the fingerprints. As the AES-256 Galois counter mode (GCM) with PIN oriented keys and lightweight lattice-based randomness generator is one of the fastest algorithms in cryptography, it is used for key tag and Fingerprint Minutiae template. This will reduce the computational cost and optimize the encryption time.

Fig. 8.

Encryption time and decryption time taken based on the fingerprints over regeneration of original template result minimal time.

Ablation analysis with other databases

Figure 9 illustrates the encryption and decryption fingerprints across different datasets, with authentication accuracy in biometric fingerprint detection. The self-analysis is performed using the FVC2002, SOCOFing dataset, and a synthetic dataset. The proposed security model achieves comparable accuracy and equivalent timing for fingerprint prediction, making it independent of other fingerprint datasets as well. The proposed model achieves a 98.69% detection accuracy on the FVC 2002 dataset, with encryption and decryption times of 0.020/0.018 s, respectively for unseen final validation. In comparison, the SOCOFing dataset achieves a 98.54% accuracy with encryption and decryption times of 0.022 s and 0.020 s, respectively, for a similar computational operation. This result justifies that the framework is not overfit to a single dataset and can equally generalise to different fingerprint data sources. This ensures the proposed model is practically suitable for real-time deployments, where consumer smart devices can process different types of fingerprint data by accurately predicting the user’s fingerprint and unknown fingerprints with high accuracy.

Fig. 9.

Performance evaluation on SOCO dataset and synthetic fingerprint for ablation study and analysis robustness under different fingerprint alteration condition.

Final output

The Fig. 10 shows the final output which shows the difference between original and matched fingerprints. Also, the proposed method easily distinguishes the unknown fingerprints with a lesser time. The consumer devices may get into attacks like spoofing attacks, replay attacks and man-in-middle attacks. But the proposed model resists these attacks using the AES encryption key generations. Because the key is not derived directly, instead the lattice transformed features combined with a hashed PIN and multi-cycle variability. This makes computationally infeasible to produce the encryption key from stolen biometric images. Since, the encryption is applied before transmission using strong symmetric AES-256 GCM blocks and the key is never shared in the raw form the man-in-middle attacks cannot retrieve plaintext biometric data.

Fig. 10.

Final Output on edge devices with secure fingerprint authentication.

The entropy of this proposed work can be derived using the encryption key. The Shannon entropy model can be considered with a KDF SHA-512 PIN hashing. This introduces a variability of lightweight LWE lattice-based randomness generator biometric inputs. The major source of this entropy is taken from the lightweight LWE lattice-based randomness-based salt, the output has 128 bits of computational entropy which overcomes the problems arises with LWE. The LWE based salt is independent of user PIN with final AES key generated KDF2 for unpredicted by assuring low entropy-based pin which cannot be easily enumerated this secures key space for fingerprint template protection based on user authentication effectively.

Result analysis with other state of art approaches

In the analysis of the proposed work, it is very important for considering the key parameters like accuracy and time consumption for preserving the integrity and privacy of the consumer data. The proposed AES-256 Galois counter mode (GCM) with PIN oriented keys and lightweight lattice-based randomness generator -based fingerprint feature transformation for key generation model, has been compared with traditional security methods and the proposed model is considerably good in terms of the security Table 1 defines the AES-256 Galois counter mode (GCM) with PIN oriented keys and lightweight lattice-based randomness generator accuracy.

Table 1.

Proposed AES-256 Galois counter mode (GCM) based LWE lattice with detection accuracy.

Cryptographic technique used in biometric security	Accuracy	Encryption time (ms)	Decryption time (ms)	Time complexity
hyperlattice-AES 12811	96.3	0.054	0.13	O log n
MCC Hashing30	88.0	0.47	0.44	O (n2)
Soliman et.al34/ Modified logistic Map	97.0	1.17	1.25	O(L2)
Chai et.al35/Probability Confidence Matrix	97.0	1.17	1.25	O(d⋅n + o⋅b⋅a + h⋅o)
Zhao et.al36 /TNCB	91.1	1.20	0.27	O (n log n)
Proposed AES GCM LWE Lattice	98.69	0.020	0.018	O (n)

From Table 1, it is proved that the proposed AES-256 Galois counter mode (GCM) Based LWE Lattice fingerprint feature transformation for key generation preserves both integrity and privacy with high accuracy and less time consumption.

Table 2 defines the Proposed security framework Advantages with traditional Security Algorithm and Table 3 defines comparison with existing state of art model with proposed. It also explains how the proposed security framework differs from conventional and traditional security systems in safeguarding consumer electronics.

Table 2.

Proposed security framework advantages with traditional security algorithm.

Attributes	Proposed AES GCM LWE lattice	Raw minutiae plaintext storage 30	Facial based Auth cancellable biometrics 34	Basic biometric with fuzzy template 36
Accuracy	98.69	88	95	96
EER	1.31	3.1	7.6	2.5
Lightweight	Yes	Yes	Yes	No
Security	Very high	Low	Medium	Plain and weak
Replay Attack resistance	Strong	Weak	Weak	Weak
Spoofing	Not Possible with 100,000 iterations	Possible	Centralized logging	Not applicable
Storage overhead	Minimal below 1 KB per record	Moderate	Moderate	High

Table 3.

Proposed work comparison with existing state of model approaches.

Reference/method	Fingerprint imposter detection accuracy
Antognazza Francesco et.al 11/kyberlattice-AES 128	96.3%
Risto Vaarandi and Hayretdin Bahsi28/LLM unsupervised Imposter Detection	F1 score = 78%
MOHD IMRAN et.al30/MCC Hashing	88%
Amit Kumar Trivedi et.al31/Delaunay triangulation of minutiae	91%
Geetika Arora et.al32/PalmHashNet	97.8%
Khan, M. S et.al33/straight- through estimation original feature	97.68%
Proposed AES GCM LWE lattice	98.69%

The generated output shows the proposed method has high MSE, low PSNR, and low SSIM compared with traditional models like ECC and RC5. The output indicates that the encrypted biometric data contains no structure and fully random with the use of lattice based random generator. Next, the decryption process occurs after creating secure access, resulting in low MSE, high PSNR, and SSIM values close to 1, which guarantees that the original fingerprint is accurately restored without any loss, making it very usable in consumer electronics and does not distort fingerprint information in between template protection process. Finally, the entropy of the fingerprint cryptography image is higher than that of a normal image, which ensures high randomness and resistance against the statistical attacks. The Fig. 11 defines image quality metrics for encrypted images on regeneration of result during encryption and decryption.

Fig. 11.

Final image quality metrics between original, decrypted fingerprint template by demonstrating reconstruction and irreversibility function.

The proposed security framework doesn’t need external servers, and it also works on devices more suitable for remote consumer electronics controlled by the particular user. The proposed lightweight security framework uses fingerprint biometrics, which are unique to each user, not static-type passwords. It also employs encryption along with verification to safeguard device security and verify credentials at runtime. The final proposed security framework has minimum latency, making it more suitable for consumer electronics like smartphones and PDA resource-constrained devices.

Conclusion and future research direction

The proposed model presents a holistic approach to securing the consumer device data. As consumer device data is constantly exposed to unauthorised risks, the AES-256 Galois/Counter Mode (GCM)-based LWE lattice-based fingerprint feature transformation for key generation provides high data integrity and privacy. The design of the encryption key enables robust authentication and precise access control. This allows the user to trust the devices from the misuse of their personal devices. Also, the proposed model achieved 98.69% accuracy, with authentication times of 0.020 and 0.018 on the given dataset, and reduced encryption and decryption time during internal validation. The model utilizes encryption and decryption functions to map the original fingerprint and analyse the spoofed fingerprint effectively. The model is better suited for deploying edge devices without requiring additional resources. Traditional systems often require extensive resources and high storage overhead to maintain secure templates. Still, they require significant resources and further training to overcome edge-deployment challenges with lattice-based cryptographic functions. The model integrates lattice-based entropy into the fingerprint key generation function, using a PIN-based KDF2 tag and an LWE salt. The proposed work achieves strong security for verifying user fingerprints with high accuracy. Experimental results with FVC2002 show strong fingerprint template accuracy, with an ablation study on the SOCO dataset demonstrating high reliability in fingerprint authentication against offline and leakage attacks.

Even though the proposed model offers improvements in data protection and integrity, this research can further enhance its robustness and applicability in the digital era. The model can be applied to quantum-resistant cryptography, AI fused ML-based data privacy, Blockchain, and the analysis of diverse datasets like livDet and WISE IoT, which would be handy in the future for resource-constrained consumer electronic devices. While focusing on these areas, the model can be more robust and better preserve consumer data. Also, it will anticipate major security threats. In the future, the model will be will be used in an innovative city environment to process user identification without the need for additional cloud deployment, and an improved transportation system will leverage this proposed security algorithm will take to utilize control devices with security against unauthorized access. The proposed work will be extended in the future to utilize other biometric functions, such as voice and facial mapping, in real-world deployments across IoT hubs, Smart city environments, and edge device security. More optimization of hardware on biometric template databases in the future direction, with a lattice-based key security function, with less computational overhead, by achieving high accuracy of results.

Author contributions

- **Aanjankumar Sureshkumar** : Conceptualization, Methodology, Data Curation, Writing—Original Draft Preparation.—**M. Maragatharajan** : Formal Analysis, Investigation, Writing—Review & Editing.—**Kaushik Jangiti** : Supervision, Validation, Project Administration, Resources.—**M. Karuppasamy** : Software Development, Model Implementation, Visualization.—**Prakash Tukaram Raut** : Statistical Analysis, Interpretation of Results, Critical Review.—**Kathiresan Jayabalan** : Experimental Design, Literature Review, Manuscript Editing.—**Nithya Rekha Sivakumar** : Data Validation, Technical Support,

Funding

Open access funding provided by Symbiosis International (Deemed University). This research is supported by Princess Nourah bint Abdulrahman University (PNU) Researchers Supporting Project number (PNURSP2026R194), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Data availability

Authors declare that all the data being used in the design and production cum layout of the manuscript is declared in the manuscript and dataset used in the research [https://www.kaggle.com/datasets/nageshsingh/fvc2002-fingerprints] and https://www.kaggle.com/datasets/ruizgara/socofing code available in [https://github.com/TECHcode317/fingersec]

Declarations

Competing interests

The authors declare no competing interests.