Generalizable deepfake detection via Spatial Kernel Selection and Halo Attention Network

The rapid advancement of AI-Generated Content (AIGC) has enabled the unprecedented synthesis of photorealistic facial images. While these technologies offer transformative potential for creative industries, they also introduce significant risks due to the malicious manipulation of visual media. Current deepfake detection methods struggle with unseen forgeries due to their inability to consider the effects of spatial receptive fields and local representation learning. To bridge these gaps, this paper proposes a Spatial Kernel Selection and Halo Attention Network (SKSHA-Net) for deepfake detection. The proposed model incorporates two key modules, namely Spatial Kernel Selection (SKS) and Halo Attention (HA). The SKS module dynamically adjusts the spatial receptive field to capture subtle artifacts indicative of forgery. The HA module focuses on the intricate relationships between neighboring pixels for local representation learning. Comparative experiments on three public datasets demonstrate that SKSHA-Net outperforms the state-of-the-art (SOTA) methods in both intra-testing and cross-testing.