System detects and explains forgery in multilingual document images
Team MSU GenText-Forensics Challenge 2026 Technical Report
Computer Vision and Pattern Recognition
Summary
Documents can be manipulated not just by changing pixels but by altering their meaning, which tricks systems that read text automatically. The authors present a system that finds whether a document’s text image is fake, points out exactly where it was changed, figures out the kind of forgery, and writes a clear explanation. Their system uses several models working together: one finds suspicious areas, a second checks these areas and guesses the attack type, and a third looks for changes in meaning that are hard to spot visually. Their method ranked third in a challenge testing such forgery detection.
What this means in practice
- •For security software developers: Detect and explain forgery in scanned or photographed documents by integrating a system that localizes changes and identifies tampering type.
- •For legal compliance teams: Use detailed forensic reports to validate document authenticity and support investigation of fraudulent documents in multilingual environments.
Authors
Kirill Koltsov, Aleksandr Gushchin, Dmitriy Vatolin, Anastasia Antsiferova
Abstract
Document text forgery has evolved beyond simple pixel-level manipulation: modern attacks alter not only the appearance of a document but also its meaning, and increasingly target the OCR & LLM pipelines that consume such documents. The ACM MM 2026 GenText-Forensics challenge therefore requires systems that not only decide whether a multilingual text image is forged, but also localize the point of manipulation, identify the attack type, and produce a human-readable forensic report with supporting evidence. We present our solution, a decomposed chain-of-thought (CoT) pipeline that combines a document tampering detector (DTD) with two Qwen3-VL-32B vision-language models, each LoRA-adapted to a distinct sub-task. DTD produces tampering probability maps that are converted into numbered candidate regions; a first model (the Filterer) validates these regions and assigns a preliminary forgery type, while a second model (the Semantic Detective) merges and re-grounds the surviving regions, searches for purely semantic anomalies that are invisible to pixel-level detectors, and writes the final report. Both models are trained by distilling chain-of-thought traces from a privileged Qwen3-VL-235B teacher that has access to ground-truth masks and reports. Our approach secured third place in the ACM MM 2026 GenText-Forensics challenge. We describe the data preparation, test-time augmentation, region rendering, distillation protocol, and training configuration in detail, and report ablations over detector thresholds, prompt designs, and pipeline decompositions.