Skip to content
Book Open access

Fine-Grained Visual and Text Forgery Distillation for Media Manipulation Detection

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 56 references

Abstract

With the rapid advancement of generative models, manipulated images and text have become increasingly prevalent, posing serious threats to information security. Existing media manipulation detection methods exhibit two key limitations: (1) they fail to capture modality-specific forgery trace. (2) they struggle to model the cross-modal interactions between tampered visual regions and manipulated text fragments. These result in degraded performance in predicting forgery types and grounding manipulated regions. To address these challenges, we propose Fiord, a fine-grained visual and text forgery distillation framework for multimodal media manipulation detection. Built upon a pretrained multimodal encoder, Fiord introduces a visual regional forgery distillation module and a text fragment forgery distillation module to capture fine-grained forgery traces. The former constructs a cross-domain discriminative space and leverages multi-scale spatial and frequency trace to capture subtle forgery patterns, while the latter identifies tampered fragments via token-level reasoning and cross-modal grounding. A unified multi-task learning framework further enables joint optimization of manipulation localization, forgery type prediction, and overall media authenticity. Extensive experiments demonstrate the effectiveness of each module, and our method outperforms the state-of-the-art methods on large-scale benchmark datasets. We will release all the source codes.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.