HFFT-Former: hybrid FFT transformer for image forgery localization

The rapid advancement of modern image editing tools has led to growing concerns about image tampering, particularly due to its potential to mislead the public and threaten privacy and security. As a result, detecting and localizing tampered regions has become a critical research challenge. Traditional methods struggle with subtle manipulations such as splicing, copy-move, and removal, which are often easier to detect in the frequency domain than in the spatial domain. Furthermore, the size imbalance between tampered regions and background areas complicates the detection task. To address these challenges, we propose an end-to-end network, HFFT-Former, which leverages frequency feature differences to precisely localize tampered regions. HFFT-Former integrates FFT and Transformer in a hybrid architecture, comprising two core modules: the Hybrid FFT Transformer (HFFT) and UP-Head. The HFFT module improves tampering feature representation by embedding both high- and low-frequency components into the self-attention mechanism via parallel dual-filter attention (DFA). The UP-Head module employs a coarse-to-fine hierarchical strategy to merge multi-scale features, thereby enhancing localization accuracy and boundary refinement. Experimental results demonstrate that HFFT-Former outperforms state-of-the-art models on three benchmark datasets and one real-world dataset, showcasing superior generalization and robustness.