Why Physics Beats AI for Deepfake Detection
The Detection Arms Race
Traditional deepfake detection focuses on identifying generator-specific artifacts: compression inconsistencies, blending boundaries, physiological implausibilities. But this approach has a fundamental problem.
As generative models improve, these artifacts disappear. We're playing defense in a game where the offense gets stronger every few months, and our tools become obsolete just as quickly.
A Different Question
Instead of asking "does this video contain artifacts?", we can ask a different question: "is this video physically consistent with the world?"
This reframing matters because physics doesn't change. The relationship between camera motion and optical flow, the timing constraints imposed by rolling shutter sensors, the spectral properties of emissive displays versus natural light: these are invariant.
Rolling shutter distortion bending a turboprop's blades. This temporal artifact is a physical constraint that synthetic video cannot easily replicate. Credit: Richmilliron, CC BY-SA 3.0
The Axiom Approach
Axiom explores this idea through multiple layers:
- Hardware binding: Cryptographic signatures tie sensor data to specific physical devices
- Physics verification: Rolling shutter timing creates dense temporal constraints that synthetic video cannot easily replicate
- Formal certification: Machine-checked proofs make assumptions and guarantees explicit
This is research in progress, not a finished product. Many open questions remain.
What's Next
Future posts will explore specific aspects of the protocol: the threat model, the physics of rolling shutter verification, and the limitations we've identified along the way.
The goal isn't to claim we've "solved" deepfake detection. It's to explore whether physics-based verification offers a more sustainable foundation than the artifact-hunting approach that dominates today.
