LAA-Net: Localized Artifact Attention Network for Quality-Agnostic and Generalizable Deepfake Detection

This paper introduces a novel approach for high-quality deepfake detectioncalled Localized Artifact Attention Network (LAA-Net). Existing methods forhigh-quality deepfake detection are mainly based on a supervised binaryclassifier coupled with an implicit attention mechanism. As a result, they donot generalize well to unseen manipulations. To handle this issue, two maincontributions are made. First, an explicit attention mechanism within amulti-task learning framework is proposed. By combining heatmap-based andself-consistency attention strategies, LAA-Net is forced to focus on a fewsmall artifact-prone vulnerable regions. Second, an Enhanced Feature PyramidNetwork (E-FPN) is proposed as a simple and effective mechanism for spreadingdiscriminative low-level features into the final feature output, with theadvantage of limiting redundancy. Experiments performed on several benchmarksshow the superiority of our approach in terms of Area Under the Curve (AUC) andAverage Precision (AP). The code is available athttps://github.com/10Ring/LAA-Net.