Practical Deepfake Detection: Vulnerabilities in Global Contexts

Authors: Yang A. Chuming, Daniel J. Wu, Ken Hong

Published: 2022-06-20 15:24:55+00:00

Comment: 6 pages, 6 figures, presented as a workshop paper at Responsible AI @ ICLR 2021

AI Summary

This paper investigates the practical efficacy of state-of-the-art deepfake detection algorithms when faced with real-world video quality degradation. By simulating various data corruption techniques on the FaceForensics++ dataset, the study reveals that while models are robust to corruptions resembling training augmentations, they remain vulnerable to significant decreases in video quality. This vulnerability can lead to misclassifying legitimate videos as deepfakes, as demonstrated with the Gabonese President Bongo's address.

Abstract

Recent advances in deep learning have enabled realistic digital alterations to videos, known as deepfakes. This technology raises important societal concerns regarding disinformation and authenticity, galvanizing the development of numerous deepfake detection algorithms. At the same time, there are significant differences between training data and in-the-wild video data, which may undermine their practical efficacy. We simulate data corruption techniques and examine the performance of a state-of-the-art deepfake detection algorithm on corrupted variants of the FaceForensics++ dataset. While deepfake detection models are robust against video corruptions that align with training-time augmentations, we find that they remain vulnerable to video corruptions that simulate decreases in video quality. Indeed, in the controversial case of the video of Gabonese President Bongo's new year address, the algorithm, which confidently authenticates the original video, judges highly corrupted variants of the video to be fake. Our work opens up both technical and ethical avenues of exploration into practical deepfake detection in global contexts.

Key findings

Deepfake detection models are robust against video corruptions that align with training-time augmentations, but remain highly vulnerable to corruptions simulating severe decreases in video quality, such as high Constant Rate Factor (CRF) and datamoshing. These vulnerabilities can lead to legitimate videos being incorrectly classified as fake, potentially causing dangerous socio-political implications in global contexts with poor infrastructure.

Approach

The authors developed a video corruption pipeline to simulate quality degradation (e.g., bitrate reduction, resolution downsampling, CRF increase, datamoshing) on videos from the FaceForensics++ dataset. They then evaluated the performance of a pre-trained, state-of-the-art deepfake detection model (an ensemble of EfficientNet B7s) on these corrupted legitimate and deepfake videos.

Datasets

FaceForensics++, Deepfake Detection Challenge Dataset (DFDC) (for model training)

Model(s)

Ensemble of 7 EfficientNet B7s

Author countries

United States

← Previous