Sample Correlation for Fingerprinting Deep Face Recognition

2024-10-25

International Journal of Computer Vision

AI Summary

Deep face recognition models are vulnerable to unauthorized copying and lack effective copyright protection and provenance tracing mechanisms. To address this, we propose a lightweight black-box fingerprinting method that uniquely identifies a target model by modeling the correlation between feature responses to input sample pairs—without modifying the model architecture or requiring retraining. Our approach leverages statistical correlation analysis and feature-space similarity metrics, ensuring both low input perturbation and high discriminability. It is rigorously validated for adversarial robustness and cross-model generalization. Evaluated on mainstream models including ArcFace and CosFace, the method achieves >99.2% fingerprint identification accuracy. It demonstrates strong resilience against common distortions such as cropping, compression, and filtering, with per-sample fingerprint extraction completed in under 0.1 seconds.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Web Mining and Content Analysis: Large pretrained models with web data

Problem

• Face Recognition
• Model Protection
• Copyright Security

Innovation

• SAC
• JPEG Compression
• Face Recognition Security