Jul 2026· International Conference on Control, Decision and Information Technologies· pp. 2711-2716· 0 citations· 18 references
Abstract
Face recognition in unconstrained environments remains a challenge due to variations of illumination, pose, expression and partial occlusion while large scale biometric deployment raises significant privacy concern with regards to processing sensitive facial data. This paper presents PrivFace-AI which is a framework for private face recognition using federated learning to securely perform biometric identification without transferring raw facial data to centralized servers. PrivFace-AI combines deep embeddings extracted from facial embedding using FaceNet; cosine similarity matching; decentralized model aggregation via FedAvg and liveness detection which improves robustness against spoofing attacks under real world operating conditions. A complete web-based implementation coupled with optional embedded hardware interaction was developed to validate practical deployment of the framework in intelligent access control scenarios. Experimental results obtained on LFW, KinectFace and custom real world datasets demonstrate 97.8% recognition accuracy in real-time response capability confirming the effectiveness of PrivFace-AI in combining privacy preservation, recognition robustness and operational feasibility under unconstrained environment.
Finger vein recognition (FVR) has significant potential in biometrics due to its high accuracy and intrinsic liveness detection capabilities. However, the increasingly stringent privacy regulations have presented severe data security challenges for traditional centralized training. While federated learning (FL) mitigates these privacy concerns through a decentralized training paradigm, conventional FL algorithms that seek a single global model experience significant performance degradation on non-independent and identically distributed (Non-IID) data in real-world cross-institutional deployments. This degradation stems primarily from a dual-heterogeneity issue that involves domain shift caused by hardware discrepancies across acquisition devices, and label skew resulting from nonoverlapping user identities. To address this dual-heterogeneity challenge, we propose a personalized federated learning framework driven by hierarchical parameter decoupling and subspace metric. First, we designed a hierarchical parameter decoupling architecture. Macroscopically, the architecture retains the classifier locally to isolate label heterogeneity; microscopically, it introduces an additive parameter decomposition that decouples the feature extractor on a global full-rank basis (to capture domain-invariant semantics, namely, the shared physiological vein topologies) and a local low-rank adapter (that accommodates device-specific characteristics, such as hardware-induced noise and illumination discrepancies). Furthermore, we propose a subspace similarity matching strategy based on principal angles on the Grassmann manifold. By exploiting the geometric properties of low-rank projection matrices, this strategy accurately quantifies the underlying distribution discrepancies among clients to guide personalized weighted aggregation. Extensive experiments on six public finger vein datasets demonstrate that the proposed framework significantly improves the overall recognition performance and mitigates performance degradation caused by data heterogeneity.
Ximing Zhou, Yuhan Wang, Jiajun Cui et al.· Italian National Conference...· 0 citations
Finger vein recognition technology has become one of the primary solutions for high-security identification systems. However, traditional finger vein recognition methods face several limitations, such as the risk of permanent identity loss due to biometric data leakage. Through designing cancelable biometrics, users' privacy and security can be further protected, and the risk of biometric data misuse can be reduced. In this paper, we propose Deep Bio-Hashing Network (DBHN) for privacy-preserving finger vein recognition, achieving end-to-end cancelable recognition. Specifically, we design a class center alignment module to improve feature alignment, which aligns the variations of all potential finger views with the finger center view via a learnable transformation. Furthermore, to tackle the security problem caused by stolen tokenized random numbers, a Deep Bio-Hashing layer is proposed, which utilizes a system-level token instead of assigning unique tokens to each user. To supervise the learning process of DBHN, we design a hybrid loss function including classification loss, consistency-based localization loss, and class center triplet loss. Finally, we conduct experiments and analysis on three publicly available datasets. Experimental results show that our method has favorable recognition performance and achieves competitive results compared to state-of-the-art hash-based methods. The analysis verifies the cancelable biometrics attributes and justifies the resilience of the method against existing security and privacy attacks.
Jie Gui, Yifan Wang, Minjing Dong et al.· IEEE Transactions on Image P...· 0 citations
The research offers an in-depth analysis of various DP techniques to construct a secure face recognition system employing a Convolutional Neural Network and face classifiers, and concludes that the DP blur with Logistic regression predictors provides the highest privacy, achieving excellent accuracy rates of 97% and 77% for these datasets.
Muhammad Minoar Hossain, Mohammad Motiur Rahman· PLoS ONE· 0 citations
Unlearnable examples keep publicly shared photos from being learned by unauthorized face-recognition models. An imperceptible perturbation, added before sharing, makes any model trained on the protected photos fail on clean faces. The perturbation is crafted on the shared image, however the attacker trains on the face it extracts, cropped and resized to the recognizer input, and under this extraction the protection collapses. We propose LPID, which builds the extraction into the unlearnable-example objective. LPID confines the perturbation to the extracted face region and optimizes it through a differentiable model of the extraction, concentrating its energy in the frequency band the extraction preserves. Because this robustness is a property of the transform rather than of any identity, LPID is re-optimized per album and protects even users it has never seen. LPID attains the lowest attacker accuracy of all methods in every setting we evaluate, holding the attacker below $10\%$ under crop+resize extraction on identities unseen at protection time, while remaining imperceptible at $32.7$\,dB PSNR and $0.161$ LPIPS.
Byunghoon Oh, Sunghwan Park, Jaewoo Lee· 0 citations
Biometric cryptosystems designed with revocability and session-level renewability must preserve template privacy and maintain reliable authentication despite the inherent variability of biometric data. A key unresolved issue in deep biometric protection is how the embedding should be formed from a network’s final representation for secure cryptographic key binding. This study presents a privacy-preserving face biometric framework in which deep embeddings are transformed into binary revocable templates and bound on the fly to cryptographic keys using a fuzzy commitment scheme with error-correcting codes. This work focuses on revocability and session-level renewability; achieving full ISO/IEC 24745 cancelability (specifically unlinkability and irreversibility) requires a keyed transform and salted-hashed storage, which we identify as essential future work. Holding the trained network and the binding pipeline fixed, three feature-aggregation strategies applied to the final representation are systematically compared, namely Global Average Pooling, Smoothed Flattening, and a dedicated Dense Feature Layer. The resulting templates are combined with Reed–Solomon and Extended Hamming coding schemes to analyse the trade-offs among discriminability, error tolerance, key length, and computational cost. Experiments conducted on the FEI Face Database show that the Dense Feature Layer (a learned non-linear projection) provides a markedly more suitable representation for secure key binding than the non-learned spatial-aggregation alternatives, achieving the best authentication performance among the tested alternatives. At the selected operating point under controlled-acquisition, frontal-face conditions, the protected system achieves a 0% false acceptance rate (95% confidence interval [0%, 0.92%], n = 400) with a 1360-bit bound key while maintaining a genuine acceptance rate of 91.0% with Reed–Solomon coding and 89.75% with Extended Hamming coding in the structurally secure independent-segment configuration, which eliminates key-reuse leakage. On the deployment side, mobile-side binding completes in well under a millisecond for both codes, whereas cloud-side decoding differs sharply: Reed–Solomon requires 15.056 ms per attempt against 0.042 ms for Extended Hamming, a≈ 350 × gap that dominates one-to-many scalability. The results further show that Reed–Solomon coding is preferable for high-security low-scale verification, whereas Extended Hamming coding offers substantially lower cloud-side decoding cost and better scalability for one-to-many identification. Consequently, the relative ranking of the three strategies, rather than the absolute error rates, is the primary transferable finding. Overall, the study provides practical design guidance for jointly selecting CNN feature representations and cryptographic binding mechanisms in secure and privacy-preserving biometric authentication systems.
Atef Bentahar, M. C. Ghanem, R. Saidi et al.· Cybersecurity· 0 citations
The proposed DiffAttack framework significantly outperforms existing adversarial techniques, achieving a high average attack success rate of 84.86% across multiple face recognition models (e.g., FaceNet).