Check Robustness of Deepfake Video Detection models with Adversarial Examples

What is it about?

The paper is about the process of attacking sequence-based deepfake detection models (Conv-LSTM and Facenet-LSTM) with adversarial attacks (FGSM and CW-L2) to craft adversarially perturbed deepfake video frames. We have successfully shown that the sequence-based deepfake video detectors can be tricked with adversarial examples.

Featured Image

Photo by Sander Sammy on Unsplash

Photo by Sander Sammy on Unsplash

Why is it important?

Prior works suggest that the sequence-based deepfake detector models are better performing than CNN-based models and are more likely to be deployed in the real world. Our experimental results strongly suggest that the sequence-based deepfake detectors are still vulnerable to adversarial attacks, both in the white-box and black-box setup. We would like to highlight the importance of developing more robust sequence-based deepfake detectors and opening up directions for future research.