Securing gait recognition with homomorphic encryption.

Area of Science:

• Biometrics and Cybersecurity
• Applied Cryptography
• Machine Learning

Background:

• Biometric identification systems offer robust security but pose privacy risks due to the inability to revoke compromised data.
• Current systems often require sensitive biometric data to be decrypted for processing, increasing vulnerability.
• The need for privacy-preserving techniques in biometric recognition is paramount.

Purpose of the Study:

• To explore the application of homomorphic encryption (HE) for protecting biometric data during the classification phase.
• To develop and evaluate a secure gait recognition system using HE.
• To analyze the impact of HE on system accuracy and computational complexity.

Main Methods:

• A system was designed with a local feature extractor (vision transformer) and an HE-compatible classifier processing encrypted data.
• The feasibility was demonstrated on a gait recognition task.
• Statistical analysis was conducted to assess accuracy and computational overhead, considering various activation functions and their polynomial approximations.

Main Results:

• The study confirmed the feasibility of using homomorphic encryption for secure and accurate gait recognition.
• The impact of HE on classification accuracy and computational complexity was quantified.
• Performance trade-offs associated with different activation functions and their approximations under HE were identified.

Conclusions:

• Homomorphic encryption is a viable technique for enhancing privacy in biometric identification systems, specifically for gait recognition.
• The proposed system offers a promising direction for secure biometric data processing.
• Careful consideration of activation functions and their approximations is necessary to balance security and computational efficiency.