Physics-Based Deepfake Detection

Overview

A deepfake detector that analyses biomechanical motion rather than visual artifacts — 478-point 3D facial landmarks per frame via MediaPipe, then chin-velocity jitter, saccade timing and micro-tremor. Vertex AI Gemini reasons over those features, and an offline physics-heuristic mode takes over with zero code changes.

What it does

• 478-point 3D facial landmark extraction per frame via MediaPipe
• Chin-velocity jitter, saccade timing, and micro-tremor as primary signals
• Event-driven pipeline on Cloud Run, Firebase Functions and Firestore
• Offline physics-heuristic mode for when cloud infrastructure isn't available