Robust Adversarial Attacks on Deepfake Detectors: A Transferable Approach

Recent developments in video manipulation techniques have lowered the barrier for creating fake videos.Manipulated videos can fuel misinformation and undermine trust in media.Hence detection for such fake videos have gained widespread interest in both academic and industrial domain.Recent DeepFake detection techniques utilize Deep Neural Networks (DNNs) to classify AI-generated fake videos from genuine videos.In this work, we illustrate one possible way to bypass such detectors: by adversarially perturbing fake videos that have been generated using existing DeepFake generation methods.We show that our adversarial perturbations are robust and successful against image and video compression codecs, meaning they present a true threat in the wild.We demonstrate pipelines that can fool DNN-based Deepfake detectors into identifying fake videos as authentic in both white-box and black-box attack scenarios.Also, using universal adversarial perturbations, which present a very realistic attack scenario due to their easy sharing among attackers, we investigate the degree to which adversarial perturbations transfer across various Deepfake detectors and develop more approachable methods.