Evaluating the Adversarial Robustness of Deepfake Detectors

As eKYC pipelines see increasingly more usage in verifying identities across banking, insurance and retail applications, bad actors are developing new ways of bypassing validation and gaining trusted access in protected environments. Deepfake detectors have their unique place in such verification pipelines and act as a defense against synthetic forgeries delivered through injection attacks. This paper evaluates the adversarial robustness of four architecturally diverse detectors under white-box and black-box attacks in a cross-domain setting. Under white-box conditions all four detectors are fully compromised, even at perturbation magnitudes that remain metrically imperceptible. Transfer attacks show that adversarial examples crafted against a single, freely available, model can evade other detectors with high reliability and query-based attacks achieve comparable results without any knowledge of model internals. These results indicate that evaluated deepfake detectors do not withstand adversarial manipulation under realistic attack conditions, raising practical concerns for production eKYC deployments and for compliance with the EU AI Act’s robustness requirement for high-risk biometric systems.