Deepfake detectors fooled by a few pixel tweaks, which is somehow still the cheap option
Ourtake
Your AI deepfake detector can be beaten by perturbations too small for a human to notice. Plan for that.
What happened
Mindgard red-teamed a deepfake detector that won a Kaggle challenge and is built on an EfficientNet B7 architecture, using three evasion attacks. SquareAttack evaded correct classification on 100% of the tested images, DeepFool on 90%, and BoundaryAttack on 30%. The company says detectors can be bypassed and recommends adversarial training, gradient masking and input modification.