Advancing Biometric Authentication: A Framework for Template Reconstruction Attack Detection

Author : Y.teja, Koutharapu Babinadh, Venkata Satya Koushik Devarabhotla, Nithin Sai Manchikalapudi, Dilip Kumar Vallabhadas

Abstract : Although fingerprint biometrics are broadly used in authentication systems, they are facing a serious problem: in case the templates are stolen or compromised, the original fingerprint patterns can be restored. Such reconstruction attacks allow adversaries to create synthetic fingerprints to bypass authentication or gain access to large-scale systems. Complementary to most traditional cancelable biometric approaches aiming at template protection, we tackle for the first time the practical problem of detecting such reconstruction attempts before causing misuse. We propose a hybrid dynamic projection-based detection framework identifying inconsistencies between genuine fingerprint embeddings and synthetically reconstructed templates. Based on a ResNet-50 backbone trained with metric learning, our system extracts robust 512-D embeddings and projects them dynamically with a hybrid projection to produce protected templates. The reconstructed samples show measurable deviations in projection consistency, score distribution, and correlation structure, allowing our system to reliably detect attacks while minimally influencing normal authentication accuracy. Based on expensive experiments using the CASIA-FingerprintV5 dataset and extensive simulated attack scenarios, we explain that the proposed approach provides strong attack detection performance at improved high recognition.

Keywords : Fingerprint biometrics, template reconstruction attack, biometric security, cancelable biometrics, dynamic random projection, deep learning, ResNet-50, attack detection, biometric forensics.

Conference Name : International Conference on AI in Data Science and Deep Learning (ICIADL-26)

Conference Place : Coimbatore, India

Conference Date : 9th May 2026