Research map / Vision-language and multimodal / Multimodal LLMs

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

ICLR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FakeShield, an explainable image forgery detection and localization framework built on a multimodal LLM (LLaVA) with a domain-tag-guided detection module and a SAM-based localization module, trained on GPT-4o-annotated MMTD-Set.

Problem

Existing image forgery detection and localization (IFDL) methods are black-box and generalize poorly across tampering types (Photoshop, DeepFake, AIGC-Editing), so they give no rationale and require users to know the tampering type in advance.

Approach

Decouple explainable detection (an M-LLM producing a verdict and textual rationale, guided by a learned domain tag that resolves data-domain conflict) from mask localization (a SAM decoder prompted by a <SEG> token from an LLM encoder that reads the generated description), with training data of image-mask-description triplets generated by GPT-4o.

Key results

• FakeShield reaches 0.98 ACC on AIGC-Editing and 0.93 on DeepFake detection, outperforming MVSS-Net and CAT-Net on the MMTD-Set.
• On localization, FakeShield achieves IoU 0.54 on CASIA1+ and 0.50 on IMD2020, ahead of OSN (0.47 and 0.38).
• Explanation CSS of 0.8860 on AIGC-Editing and 0.8446 on DeepFake, versus 0.6751 for the best baseline M-LLM (InternVL2-26B) on AIGC-Editing.

Measured effects reported in the paper

Surprising findings

• Removing the domain tag generator lowers IMD2020 ACC by 0.12, showing domain conflict is a major issue in joint multi-domain training.
• Pre-trained M-LLMs often flag real images as tampered, and only FakeShield correctly judges the real IMD2020 sample shown in the appendix.

Limitations

• Suboptimal on complex deepfake manipulations such as identity switching and full-face generation.
• Localization IoU remains low on some datasets (e.g., NIST 0.32, Korus 0.17, DeepFake 0.14), and error correction in MFLM did not help.

Experimental setup

Datasets: CASIAv2 Fantastic Reality FFHQ FaceAPP COCO CASIAv1+ Columbia IMD2020 Coverage NIST DSO Korus

Benchmarks: CASIA1+ IMD2020 Columbia Coverage NIST DSO Korus DeepFake (FaceAPP/FFHQ) DFFD Seq-DeepFake ControlNet inpainting SDXL inpainting

Metrics: ACC F1 IoU CSS (cosine semantic similarity)

Baselines: SPAN ManTraNet HiFi-Net PSCC-Net CAT-Net MVSS-Net OSN CADDM HiFi-DeepFake RECCE Exposing GPT-4o

Builds on: LLaVA LISA SAM MVSS-Net LoRA

Closest papers by meaning

• Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.91)
• ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.89)
• Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 2025 · similarity 0.89)
• Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.88)
• Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.88)
• SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.88)