Image Forgery Detection Using Attention-Aware Hierarchical-Feature Fusion Module Approach

Abstract

Image forgery detection is a crucial task in digital forensics, aiming to identify manipulated or tampered images. Traditional approaches commonly use handcrafted features, which struggle to effectively capture complex patterns of forgeries. Deep learning techniques have shown promise in this domain, but leveraging hierarchical features while dynamically attending to relevant information remains a challenge. The Attention-aware Hierarchical-feature Fusion Module (AHFM) is proposed to address this issue. The module uses attention mechanisms to selectively fuse hierarchical features extracted from different levels of a convolutional neural network (CNN), enhancing the discriminative power of the network for forgery detection tasks. Experimental results show the AHFM achieves state-of-the-art performance in accuracy and robustness against various forgery techniques. Qualitative analyses provide insights into the interpretability and efficacy of the attention mechanism in identifying forged regions within images. We conduct experiments on a typical dataset of CASIA v1.0 and v2.0 fabricated images, both before and after processing, to illustrate the theoretical concept of the suggested method. Furthermore, we relate the results to those of contemporary methods in order to demonstrate our superior detection rates.