A Tutorial on Hacking Facial Recognition Systems

### Tutorial on Evaluating Facial Recognition Security At 00:00, Richard Harang, Director of Data Science at Sophos, and Dr. Ethan Rudd deliver a tutorial evaluating the security and adversarial attack surfaces of facial recognition systems at the AI Village.

### Deep Metric Learning and Feature Space Embeddings At 06:17, Harang breaks down the architecture of modern face recognition models (such as FaceNet and VGG-Face). Rather than simple classification, these networks utilize deep metric learning to project face images into high-dimensional embedding spaces where Euclidean distance defines identity similarity.

### Dataset Skew and Algorithmic Bias At 12:31, the speakers address demographic bias in computer vision. They explain how skews in public training datasets generate uneven error rates across racial and gender demographics, creating exploitable verification vulnerabilities.

### Gradient-Based Adversarial Evasion Attacks At 18:57, the tutorial reviews mathematical mechanics for crafting adversarial perturbations. Utilizing Projected Gradient Descent (PGD), attackers compute gradients with respect to input pixels, creating imperceptible noise patterns that push embeddings across decision boundaries to cause false matches or evasions.

### Physical Eyeglass Frames and Defensive Countermeasures Concluding between 25:43 and 33:28, the presenters explore real-world physical attacks, including printed adversarial eyeglass frames designed to fool surveillance systems. They evaluate defensive countermeasures, including 3D liveness detection, thermal imaging, and adversarial retraining.

(AI summary — Gemini Flash)