Learning Universal Features for Generalizable Image Forgery Localization

TL;DR

This work addresses the challenge of detecting and localizing unseen image forgeries by proposing GIFL, a framework that learns universal authentic content features rather than forgery-specific traces. It introduces a frozen encoder-driven pipeline with authentic-feature reconstruction and a Universal Forgery Localization Transformer (UFLT) that fuses spectral and spatial cues to align forged and pristine regions, enabling robust generalization. To support research in modern deep-generative forgeries, the authors create Forgery ADE, a large, diverse dataset consisting of eight forgery methods applied to ADE 20K, with authentic negatives and irregular masks. Experimental results show GIFL achieves state-of-the-art performance on unseen forgeries and competitive results on seen ones, while providing extensive analyses on data-related factors and practical training configurations. The work also shares code and dataset to facilitate broader adoption and further study of universal forgery localization.

Abstract

In recent years, advanced image editing and generation methods have rapidly evolved, making detecting and locating forged image content increasingly challenging. Most existing image forgery detection methods rely on identifying the edited traces left in the image. However, because the traces of different forgeries are distinct, these methods can identify familiar forgeries included in the training data but struggle to handle unseen ones. In response, we present an approach for Generalizable Image Forgery Localization (GIFL). Once trained, our model can detect both seen and unseen forgeries, providing a more practical and efficient solution to counter false information in the era of generative AI. Our method focuses on learning general features from the pristine content rather than traces of specific forgeries, which are relatively consistent across different types of forgeries and therefore can be used as universal features to locate unseen forgeries. Additionally, as existing image forgery datasets are still dominated by traditional hand-crafted forgeries, we construct a new dataset consisting of images edited by various popular deep generative image editing methods to further encourage research in detecting images manipulated by deep generative models. Extensive experimental results show that the proposed approach outperforms state-of-the-art methods in the detection of unseen forgeries and also demonstrates competitive results for seen forgeries. The code and dataset are available at https://github.com/ZhaoHengrun/GIFL.

Figures (7)

• Figure 1: Illustration of the traditional classification-based image forgery detection pipeline (left) and our proposed GIFL method (right).
• Figure 2: Architecture of our proposed Universal Forgery Localization Transformer (UFLT) (red part), which contains multiple dual-domain attention components (cyan part).
• Figure 3: Forged images generated by different forgery methods in Forgery ADE dataset.
• Figure 4: Visual comparison of our results and those of previous methods. Red boxes indicate the results on seen forgeries and blue for unseen ones. Zoom-in to see the details.
• Figure 5: Visualization of target features and reconstruction features of different types of forged images. Zoom-in to see the details.