DF2023-XAI: Auditing Image Forgery Localization via Scene-Disjoint Evaluation and XAI Faithfulness

This repository contains the official implementation of DF2023-XAI, a reproducible evaluation framework for image forgery localization designed to address metric inflation and spurious learning in forensic deep learning models. The framework enforces a strict Scene-Disjoint evaluation protocol to eliminate background leakage and supervisory conflict, and introduces Saliency Total Variation (TV) as a quantitative metric for auditing attribution faithfulness in explainable AI (XAI). The codebase supports comparative benchmarking of convolutional and Transformer-based architectures (e.g., U-Net and SegFormer) under controlled conditions, enabling rigorous analysis of the accuracy–faithfulness trade-off in modern forgery localization systems. It includes dataset partitioning utilities, training and evaluation pipelines, XAI attribution generation (Grad-CAM++ and Integrated Gradients), ablation experiments, and fully reproducible configurations.