SIDWA: Synthetic Image Detection Based on Discrete Wavelet Transform Stem and Deformable Sliding Window Cross-Attention

With the rapid evolution of Generative Adversarial Networks (GANs) and diffusion models (DMs), the detection of synthetic images faces significant challenges due to non-rigid artifacts and complex frequency biases. In this paper, we propose SIDWA, a novel dual-branch detection framework that leverages the synergy between frequency and spatial domains. Within the spatial branch, we design a Deformable Sliding Window Cross-Attention (DSWA) module, which utilizes a learnable offset mechanism to dynamically warp the receptive field, effectively capturing distorted edges and non-linear texture features. Simultaneously, the Discrete Wavelet Transform (DWT) Stem decomposes input images into multi-scale sub-bands to preserve crucial high-frequency residues. Through a Frequency-Semantic Resonance Projector (FSRP) strategy, the semantic priors from the spatial branch act as queries to guide the model toward localized frequency anomalies, achieving a unified “where to look” and “how to analyze” approach. Experimental results for the SIDataset (SIDset) benchmark demonstrate that Synthetic Image Detection based on Discrete Wavelet Transform Stem and Deformable Sliding Window Cross-Attention (SIDWA) achieves superior performance, with an average accuracy exceeding 95% and a competitive inference time of 18.2 ms on an NVIDIA A100 GPU. Ablation studies further validate the critical role of learnable offsets and frequency integration in enhancing robustness and generalization. SIDWA offers an efficient and reliable forensic solution for combating the growing threats of sophisticated generative forgeries.