Face Recognition¶
Face recognition systems automatically identify or authenticate individuals from facial images or video footage using neural networks or classical computer vision methods. Such systems are increasingly deployed in security, law enforcement, and identity verification contexts. However, they are vulnerable to both spoofing attacks (presentation attacks) and deep synthesis attacks like deepfakes.
Key challenges¶
Spoofing and presentation attacks: Simple attacks using printed photos, masks, or display-based replay can fool recognition systems.
Adversarial robustness: Subtle perturbations to images or deepfake manipulation can cause dramatic failure in state-of-the-art systems, with false acceptance rates exceeding 95%.
Fairness and demographic bias: Recognition systems exhibit higher error rates on certain demographic groups (darker skin tones, women, underrepresented ethnicities), limiting deployment fairness.
Cross-domain generalization: Models trained on one set of recording conditions fail on video from different lighting, pose, or resolution.
Detection and security¶
Because deepfakes are increasingly realistic, traditional liveness detection (determining whether a face is real or video-based) requires multi-modal approaches:
- Frequency-domain analysis: Looking for compression artifacts from GAN synthesis
- Temporal consistency checks: Verifying facial expressions, eye blinks, and micro-expressions are naturalistic
- Audio-visual synchronization: Detecting lip-sync mismatches or timing inconsistencies
- Challenge-response protocols: Asking subjects to follow head movements or perform actions
Related topics¶
- Deepfakes — synthetic videos that can fool recognition systems
- Synthetic Media Detection — technical methods for detecting manipulation
- Presentation Attacks — spoofing and anti-spoofing in biometric systems
Key papers in this wiki¶
- DeepFakes: a New Threat to Face Recognition? Assessment and Detection — demonstrates VGG and FaceNet achieve 85.62% and 95.00% FAR on deepfakes