Physics-Inspired Deepfake Detection via Hamiltonian Dynamics
A new approach to deepfake detection, Hamiltonian Action Anomaly Detection (HAAD), is proposed in arXiv:2605.04405. The method shifts from static pattern recognition to dynamical stability analysis, hypothesizing that natural images occupy stable low-energy states while deepfakes are in unstable high-energy states. The framework models the image latent manifold as a potential energy surface and uses Hamiltonian dynamics to detect anomalies. This physics-inspired method aims to break the cycle of periodic recalibration required by current detectors as generative models evolve.
Key facts
- arXiv:2605.04405 proposes HAAD for deepfake detection
- HAAD uses Hamiltonian dynamics and potential energy surfaces
- Natural images are hypothesized to be in low-energy stable states
- Deepfakes are hypothesized to be in high-energy unstable states
- The method aims to avoid periodic recalibration of detectors
- It is motivated by physics-inspired priors
- The approach moves from static pattern recognition to dynamical stability analysis
- Generative models do not enforce geometric smoothness constraints
Entities
Institutions
- arXiv