Fighting against deepfakes in the wild

Abstract

Deepfakes are fake media generated by deep learning models. Deepfakes can easily give attackers the ability to control one's identity. Hence, attackers can make use of deepfakes to achieve their malicious purposes such as defamation and spreading misinformation.

As deepfake generation tools become more and more readily available, the threat posed by deepfakes looms large. Therefore, it is crucial to develop new ideas to detect deepfakes. The Trusted Media Challenge organized by AI Singapore has given us the chance to explore deepfake detection methods. By participating in this challenge, our team has had the opportunity to attempt to solve fake face detection, fake voice detection and inconsistency detection. This report aims to summarize the models and techniques used for each kind of deepfake detection.