Interpretability of Deep High-Frequency Residuals: A Case Study on SAR Splicing Localization

Multimedia Forensics (MMF) investigates techniques to automatically assess the integrity of multimedia content, e.g., images, videos, or audio clips. Data-driven methodologies like Neural Networks (NNs) represent the state of the art in the field. Despite their efficacy, NNs are often considered "black boxes" due to their lack of transparency, which limits their usage in critical applications. In this work, we assess the interpretability properties of Deep High-Frequency Residuals (DHFRs), i.e., noise residuals extracted from images by NNs for forensic purposes, that nowadays represent a powerful tool for image splicing localization. Our research demonstrates that DHFRs not only serve as a visual aid in identifying manipulated regions in the image but also reveal the nature of the editing techniques applied to tamper with the sample under analysis. Through extensive experimentation on spliced amplitude Synthetic Aperture Radar (SAR) images, we establish a correlation between the appearance of the DHFRs in the tampered-with zones and their high-frequency energy content. Our findings suggest that, despite the deep learning nature of DHFRs, they possess significant interpretability properties, encouraging further exploration in other forensic applications.