FSDNet: A Decoupled Forgery Style Network for Image Forgery Detection

In the current digital era, the rapid advancement of image editing technologies has made synthetic image generation increasingly accessible, posing significant threats to digital content security and authenticity. To address this challenge, this paper proposes a novel framework called Forgery Style Decoupling Network (FSDNet), which employs a multi-point approach for effective forgery detection and style disentanglement. First, we utilize a contrastive enhancement network to amplify feature discrepancies between forged and authentic images. Through a series of convolutional layers, the network extracts and compares low-level features from both types of images, establishing a foundation for subsequent style separation. Second, a dual-path output network performs style disentanglement on the enhanced images, effectively isolating synthetic styles from authentic styles within each image. This enables more precise identification of forged content. Furthermore, we introduce a causal discriminator that achieves high-precision forgery identification by training on the dual-style representations extracted from the network outputs. Experimental results demonstrate FSDNet's superior performance, achieving 91.1% accuracy for facial images and 84.9% for landscape/object images, significantly outperforming other state-of-the-art models. These findings validate the proposed method's exceptional capability in forgery detection and its substantial application potential.