1. The Trap of Supervised Classifiers
When security researchers train a classifier on specific attack samples, the model often overfits to quirks of the attack generator rather than learning genuine human behavior.
[!NOTE] Attackers evolve their generator parameters constantly. A defense that models specific attack signatures becomes obsolete the moment the attacker modifies their retiming script.
2. The Self-Supervised Paradigm
In audio-visual forensics:
- Chugh et al. (ACM MM 2020 - Modality Dissonance Score): Evaluates whether audio track phonemes synchronize with video lip movements.
- Feng et al. (CVPR 2023): Proved that synchrony can be learned exclusively from unmanipulated genuine videos without ever seeing a deepfake.
3. Self-Supervised Calibration in CMCC
CORROBORATE adapts this methodology to web dynamics:
- The CMCC engine is calibrated solely against genuine human sessions ($A_0$).
- Per-window coherence is scored against normal human variance.
- When coherence falls below a calibrated rolling threshold for 2+ consecutive windows, a seam is flagged.