New Face, Who Dis? Recent Adversarial Approaches to Facial Recognition
summary
Survey of recent adversarial methods compares concrete tools and measured results: pixel-level perturbations (early, fragile), GAN avatars, the Camera Adversaria style procedural-noise overlay, and the cloaked-poisoning systems Fox and Loki (Fox takes ~2–5 minutes per image; Loki runs as a hosted web service that emails results and was more robust in published tests). A newer one person one mask method yields roughly 33–66% misidentification and can run near real-time with PyTorch; an open-source mobile proof-of-concept applies fast style transfer plus procedural noise, detects faces locally, and stores protected images for later upload.
AI-written summary · about
chapters (5)
- 0:00Start
- 9:18Generative Adversarial Networks
- 12:02Cloaked Poisoning: Fawkes and Lowkey Original
- 15:10Effectiveness of Fawkes and Lowkey
- 16:44Cloaking: One Person One Mask OPOM
related talks
40mNew Face, Who Dis? Protecting Privacy in a World of Surveillance
28mFace spoofing detection: theory and practice
08mAdvancements in facial recognition
57mFoundations of Facial Recognition
26mVirtual U: Defeating Face Liveness Detection by Building Virtual Models...
52mFace Recognition: From Scratch To Hatch
31mFace Recognition
24mChallenges in Real-Life Face Recognition