Skip to content
Book Open access

Towards Explainable Tampered Text Detection via Semantic-Guided Multi-Evidence Localization

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 38 references

Abstract

Tampered text detection is increasingly important for digital forensics, media integrity assessment, and document security, since manipulated text can substantially alter image semantics while remaining visually plausible. However, existing methods remain insufficient in two key aspects. First, they provide limited evidence-grounded interpretability, making it difficult to relate predictions to forged text content, low-level visual artifacts, and high-level semantic inconsistencies. Second, they rely primarily on visual appearance and therefore struggle to capture subtle forensic traces and precisely delineate small, elongated, and irregular tampered text regions. To address these challenges, we propose SMELT, a Semantic-guided Multi-Evidence Localization framework for explainable Tampered text detection. SMELT consists of a semantic-guided evidence prompting branch and a multi-evidence localizer. The former extracts tamper-relevant semantic cues from a vision-language model and converts them into localization prompts. Subsequently, the multi-evidence localizer jointly exploits visual appearance, compression-aware frequency evidence, and boundary-aware high-resolution decoding to produce precise tamper masks. To facilitate this task, we further construct VISTA-Text, an interpretable tampered text dataset with diverse domains, tampering types, and structured annotations covering authenticity judgment, manipulated text content, low-level visual artifacts, and high-level semantic inconsistencies. Extensive experiments on VISTA-Text and additional benchmarks demonstrate that SMELT consistently improves tampered text localization performance while supporting evidence-grounded interpretation.

Read PDF

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