Secure Knuckle Print Authentication: Template Protection and Attack Analysis

Biometric authentication using finger knuckle patterns offers a secure and forgery-resistant alternative for identity verification. Unlike fingerprints or facial recognition, knuckle prints are rarely exposed in daily activities, reducing the risk of unauthorized acquisition and spoofing. This study presents a privacy-preserving knuckle print authentication system using a fine-tuned DenseNet-121 model for feature extraction and homomorphic encryption for secure template matching. The IIT Delhi Finger Knuckle Database (version 1.0), consisting of 790 images from 158 individuals, is used. Data augmentation techniques, including rotation, shifting, zooming, shear transformations, and flipping, expanded the training dataset from 632 to 3,160 images, improving generalization. Feature extraction is performed using DenseNet-121, followed by homomorphic encryption of feature templates. Matching is conducted using Euclidean distance on encrypted vectors, ensuring privacy without compromising accuracy. At the optimal authentication threshold of 47, the system achieves 93.67% accuracy, with a False Acceptance Rate (FAR) of 1.099%, False Rejection Rate (FRR) of 6.329%, True Acceptance Rate (TAR) of 93.67%, and True Rejection Rate (TRR) of 98.901%. Euclidean distance outperforms cosine similarity in balancing accuracy and computational efficiency. A security analysis confirms resilience against replay attacks, feature spoofing, and template inversion through nonce-based authentication, homomorphic encryption, and secure key management. The system adheres to ISO/IEC 24745:2022 biometric security standards, ensuring irreversibility, unlinkability, and renewability. The proposed system is well-suited for high-security applications such as banking, healthcare, access control and forensic identification.