FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models

TL;DR

This work tackles the problem of reliable image forgery detection and localization across diverse tampering types by introducing FakeShield, a two-branch, multi-modal framework that leverages a domain-tagged explainable detector and a tamper-focused localization module guided by an M-LLM. It builds the GPT-4o–augmented MMTD-Set to train the system, enabling both pixel-level localization and textual explanations of tampering rationales. Empirical results show superior detection accuracy, precise localization, and richer explanations across Photoshop, DeepFake, and AIGC-editing tampering, with robustness to common degradations. By providing explainable, cross-domain forgery analysis, FakeShield advances practical digital-forensics workflows and supports applications in misinformation mitigation and legal evidence gathering.

Abstract

The rapid development of generative AI is a double-edged sword, which not only facilitates content creation but also makes image manipulation easier and more difficult to detect. Although current image forgery detection and localization (IFDL) methods are generally effective, they tend to face two challenges: \textbf{1)} black-box nature with unknown detection principle, \textbf{2)} limited generalization across diverse tampering methods (e.g., Photoshop, DeepFake, AIGC-Editing). To address these issues, we propose the explainable IFDL task and design FakeShield, a multi-modal framework capable of evaluating image authenticity, generating tampered region masks, and providing a judgment basis based on pixel-level and image-level tampering clues. Additionally, we leverage GPT-4o to enhance existing IFDL datasets, creating the Multi-Modal Tamper Description dataSet (MMTD-Set) for training FakeShield's tampering analysis capabilities. Meanwhile, we incorporate a Domain Tag-guided Explainable Forgery Detection Module (DTE-FDM) and a Multi-modal Forgery Localization Module (MFLM) to address various types of tamper detection interpretation and achieve forgery localization guided by detailed textual descriptions. Extensive experiments demonstrate that FakeShield effectively detects and localizes various tampering techniques, offering an explainable and superior solution compared to previous IFDL methods. The code is available at https://github.com/zhipeixu/FakeShield.

Figures (19)

• Figure 1: Illustration of the conventional IFDL and explainable IFDL framework. Conventional methods offer only detection results and tampered masks. We extend this into a multi-modal framework, enabling detailed explanations and conversational interactions for a deeper analysis.
• Figure 2: Illustration of the construction process of our MMTD-Set. We sample the tampered image-mask pairs from PS, DeepFake, and AIGC benchmarks, and then use domain tags to guide GPT-4o in constructing the judgment basis and focusing on both pixel-level details and image-level content.
• Figure 3: The pipeline of FakeShield. Given an image $\mathbf{I}_{ori}$ for detection, it is first processed by the Domain Tag Generator $\mathcal{G}_{dt}$ to obtain a data domain tag $\mathbf{T}_{tag}$. The tag $\mathbf{T}_{tag}$, along with the text instruction $\mathbf{T}_{ins}$ and image tokens $\mathbf{T}_{img}$, are simultaneously input into the fine-tuned LLM, generating tamper detection result and explanation $\mathbf{O}_{det}$. Subsequently, $\mathbf{O}_{det}$ and $\mathbf{T}_{img}$ are input into the Tamper Comprehension Module $\mathcal{C}_{t}$, and the last-layer embedding for the <SEG> token $\mathbf{h}_{\texttt{<SEG>}}$ serves as a prompt for SAM, guiding it to generate the tamper area mask $\mathbf{M}_{loc}$.
• Figure 4: Detection, localization and explanation results of our FakeShield.
• Figure 5: Comparisons between our FakeShield and other competitive methods. The samples, from left to right, are drawn from IMD2020, CASIA1+, Columbia, NIST16, Korus, DSO, and DeepFake.