Enhancing biometric authentication privacy and security: A synergistic approach using cancelable biometrics and federated learning

While biometric authentication offers higher security than other methods, the compromise of biometric data generates major privacy concerns. Combining federated learning for distributed privacy protection with non-invertible transformations and deep learning-based feature extraction, this work proposes a novel cancelable biometric architecture. The approach uses pretrained CNNs-MobileNetV3Small and ResNet50V2-to extract features from intermediate layers and random projection and kernel PCA are used to generate irreversible biometric templates. Secure model training guaranteed by federated learning protects raw biometric data. Using MobileNetV3Small features from layers-7 and-8, experimental results on three benchmark datasets-AMI (ear), ORL (facial), and IITD (iris)-showcase 100% or near-perfect accuracy for KNN classifiers. Using layer-7 features, the SVM on the AMI dataset attained an F1-score of 0.9665 and an accuracy of 97.8%. The proposed transformation pipeline improves accuracy by 9.16% over baseline approaches without proposed method. These findings confirm that federated learning preserves privacy without compromising recognition efficiency and that mid-level CNN features provide improved discrimination. This work proposes a deployable cancelable biometric solution concurrently addressing accuracy, revocability, and distributed security in modern authentication systems.