KSCAMa: A Lightweight Mamba Network Enhanced with Dual Attention for Image Tampering Localization

The rapid advancement of image-editing techniques presents significant challenges to the reliability of digital media. Existing image-tampering detection methods rely predominantly on convolutional neural networks (CNNs) and Transformers. However, CNNs struggle to effectively utilize contextual cues due to their fixed receptive fields, and the adoption of transformers as foundational architectures entails a significant computational burden. Recently, visual state-space models (SSMs) have emerged as a promising alternative in computer vision, demonstrating the ability to capture global dependencies with linear complexity. In this paper we introduce KSCAMa, a lightweight Mamba-based network augmented with attention mechanisms for precise image-tampering localization. The framework employs an atrous scanning strategy for multi-angle state-space modeling, integrates Kernel-Selective Fusion Attention to leverage global feature interactions, and incorporates Convolutional Additive Self-Attention in the decoder to enhance feature fusion. Experiments across eight public datasets confirm that KSCAMa achieves state-of-the-art detection accuracy and generalization performance while maintaining a competitive computational profile.