Publication

Abstract : Biometric authentication offers convenience but suffers from a major limitation: biometric traits are permanent and cannot be replaced if compromised. Cancelable biometrics addresses this challenge by transforming biometric data into non-invertible, revocable templates. This study develops a deep learning-based framework for cancelable fingerprint template generation. Features are extracted from multiple convolutional neural network (CNN) backbones including ResNet18, ResNet50, MobileNetV2, VGG16, and DenseNet121. These embeddings are transformed using three secure techniques: Random Projection (RP), Gaussian Random Projection (GRP) with binary quantization, and Biohashing. Templates are matched using Hamming distance to ensure non-invertibility and revocability. Experimental results indicate that Biohashing achieved an Equal Error Rate (EER) of 15.9%, True Acceptance Rate (TAR) of 95.8% at 1% False Acceptance Rate (FAR), and revocability of 0.49, while GRP and RP provided competitive trade-offs. The framework illustrates that integrating deep CNN embeddings with lightweight transformation techniques can deliver secure, accurate, and revocable biometric templates, ensuring both robust authentication and template protection.