Learnable Local Similarity for Face Forgery Detection and Localization

Abstract

The emergence of many face forgery technologies has led to the widespread of forgery faces on the Internet, causing a series of serious social impacts, thus face forgery detection technology has attracted increasing attention. While many face forgery detection algorithms have demonstrated impressive performance against known manipulation methods, their efficacy tends to diminish severely when applied to unknown forgeries. Previous research commonly viewed face forgery detection as a binary classification problem, disregarding the crucial distinction between real and forged faces, thereby limiting the generalizability of detection algorithms. To overcome this issue, this paper proposes a novel face forgery detection method that utilizes a trainable metric to learn local similarity between local features of facial images, achieving a more generalized detection result. What's more, it incorporate cross-level features to accurately locate forgery regions. After conducting extensive experiments on FaceForensics++, Celeb-DF-v2, and DFD, which demonstrate that the effectiveness of the proposed method is comparable to state-of-the-art detection algorithms.