Abstract
Because biometric templates kept in databases can be stolen and misused, Face Recognition technologies now have significant security vulnerabilities. Accurate recognition is often compromised by traditional template protection techniques. Strong solutions that maintain high efficiency while guaranteeing template security are required due to the increasing use of face authentication. This study proposes a deep learning-based architecture that uses InsightFace and ResNet100 to extract features while safeguarding cryptographic templates. To protect biometric data, the system employs revocable transformations, random projection, and dynamic salting. L2 normalization standardizes feature magnitudes, enabling consistent similarity comparison while maintaining the irreversibility of transformed templates. After MTCNN based face identification, the pipeline extracts 512-dimensional embeddings and performs L2 normalization. AES-256 encryption and XOR-based revocation are used to protect the features after they are projected into a 400-dimensional space and binarized using a median threshold. The protected templates are matched using cosine similarity. With EERs of 0.00% and 6.80%, respectively, the evaluation reveals with VGGFace2 and LFW datasets 99.98% and 93. 23%, respectively. The method outperforms BioHashing with an accuracy of 11.5% in VGGFace2 and offers reversible templates. It achieves 400-bit security with less than 2% FRR in a spoken dataset, which is very viable in the real world and offers a better balance between security and performance than existing methods.

This publication has 19 references indexed in Scilit: