ML Security PapersFOCA: Frequency-Oriented Cross-Domain Forgery Detection, Localization and Explanation via Multi-Modal Large Language Model
Zhou Liu 1,2, Tonghua Su 1,3,4, Hongshi Zhang 1, Fuxiang Yang 1, Donglin Di 1,2, Yang Song 5, Lei Fan 2,5
1 Harbin Institute of Technology
3 Guangdong Laboratory of Artificial Intelligence and Digital Economy
Published on arXiv
2602.18880
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
FOCA outperforms state-of-the-art IFDL methods (including MVSS-Net, HiFi-Net, SIDA, ForgeryGPT) in both detection accuracy and interpretability across spatial and frequency domains on the introduced FSE-Set benchmark.
FOCA
Novel technique introduced
Advances in image tampering techniques, particularly generative models, pose significant challenges to media verification, digital forensics, and public trust. Existing image forgery detection and localization (IFDL) methods suffer from two key limitations: over-reliance on semantic content while neglecting textural cues, and limited interpretability of subtle low-level tampering traces. To address these issues, we propose FOCA, a multimodal large language model-based framework that integrates discriminative features from both the RGB spatial and frequency domains via a cross-attention fusion module. This design enables accurate forgery detection and localization while providing explicit, human-interpretable cross-domain explanations. We further introduce FSE-Set, a large-scale dataset with diverse authentic and tampered images, pixel-level masks, and dual-domain annotations. Extensive experiments show that FOCA outperforms state-of-the-art methods in detection performance and interpretability across both spatial and frequency domains.
Key Contributions
- FOCA: first MLLM-based framework fusing RGB spatial and frequency-domain features via a Frequency Attention Fusion (FAF) module for interpretable forgery detection and pixel-level localization
- FSE-Set: large-scale dataset of 100,000 images with pixel-level tampering masks and dual-domain (RGB + frequency) natural language annotations for training and evaluating explainable IFDL systems
- Human-interpretable cross-domain explanations of tampering artifacts via MLLM reasoning over both spatial and wavelet-frequency cues
🛡️ Threat Analysis
FOCA is a detection system for AI-generated/manipulated image content — it identifies and localizes tampered regions produced by generative models. ML09 explicitly covers AI-generated content detection (deepfakes, synthetic image detection). The paper introduces both a detection framework and a dataset (FSE-Set) for evaluating image forgery authenticity, which is a content integrity and provenance problem.