网络与信息安全学报 (Apr 2024)
Robust deepfake detection method based on siamese network
Abstract
The proliferation of deepfake (DF) technology for generating manipulated facial expressions in synthetic images has raised concerns due to its potential negative impacts on individuals and society. In response to the need for robust detection, researchers have been developing methods to identify deepfakes. While current detection methods perform well on high-quality images, they often falter when confronted with low-quality or compressed images. This study focused on enhancing the robustness of deepfake detection methods to address these limitations. A novel approach leveraging a Siamese network was proposed, designed to learn common forgery features across both high-quality and low-quality images. This was achieved by trading off some of the high-quality image feature extraction capabilities to bolster the representational capacity for low-quality images. The proposed method demonstrated an average accuracy exceeding 90% across various datasets with different compression levels, surpassing several existing detection techniques. The simplicity, effectiveness, and adaptability of the proposed method to different backbone networks were further substantiated through ablation experiments.
