A new framework for attacking deepfake detectors
The MIG-COW framework uses gradient decomposition to create adversarial examples for deepfake detectors.
Researchers proposed MIG-COW to create adversarial examples. These examples are visually imperceptible and highly generalizable. They target deepfake detectors in diverse settings.
MIG-COW combines momentum-integrated gradients with a consensus-orthogonal decomposition. This method captures shared and model-specific vulnerabilities. It works across heterogeneous CNN and ViT detectors.
On the AADD-2025 Challenge benchmarks, MIG-COW reaches a 99.96% white-box attack success rate (ASR). It also maintains high perceptual similarity (SSIM).
The framework achieved a 7.16% ASR against official black-box targets. This highlights the ongoing challenge of transferability. Including low-performing but diverse models in the ensemble can degrade attack effectiveness.
Careful surrogate model selection is needed in real-world adversarial settings. This is because diverse models can lower the effectiveness of the attack.
Key facts
- MIG-COW achieved a 99.96% white-box attack success rate (ASR) on AADD-2025 Challenge benchmarks.
- The attack success rate (ASR) against official black-box targets was 7.16%.
- MIG-COW uses momentum-integrated gradients and consensus-orthogonal decomposition.