Benchmarking Deepfake Attacks on Deep Face Recognition Systems

Researchers created a testing framework to evaluate how deepfakes (AI-generated fake videos or images of people) can fool face recognition systems (AI that identifies people by their faces). The study found that deepfake attacks succeed over 70% of the time, sometimes exceeding 90%, and discovered that attack success depends more on how well attackers can control the person's identity in the fake content rather than on how realistic the deepfake looks visually.