Lightweight End-to-End Patch-Based Self-Attention Network for Robust Image Forgery Detection

Image tampering detection and localization have become critical tasks in digital forensics due to the widespread accessibility of powerful image editing tools. We came up with a new Vision Transformer (ViT)-based design that uses patch embedding and multi-head self-attention mechanisms to efficiently find fake documents. This is because we need lightweight but highly accurate detection systems that can model all of a document’s features. Our model uses a modular pipeline made up of a patch embedding layer, four transformer encoder blocks, and a streamlined classification head. The slimmed-down classification head consisted of only 69,793 trainable parameters to keep the computation costs as minimal as possible. Experimental evaluation conducted on four benchmark datasets—CASIA v1, DEFACTO (Splicing), MICC-F2000, and Columbia—demonstrates the superior performance of the proposed model. The model gets an AUC of 0.9712 and an F1 score of 0.9587 on CASIA v1, an almost perfect AUC of 0.9993 and an F1 score of 0.9914 on DEFACTO, a good AUC of 0.8491 and an F1 score of 0.8319 on Columbia, and an exeptional AUC of 0.9804 and an F1 score of 0.9097 on MICC-F2000. The proposed architecture was evaluated against 59 different models and demonstrated its effectiveness, showing its strong potential for practical applications such as detecting and localizing document forgeries.