GenText-Forensics: Challenge on Explainable Forensics and Adversarial Generation for Text-Centric Images

Fanwei Zeng, Jianshu Li, Changtao Miao, Chenqi Kong, Youru Li, Zhi Cai, Haiyan Yin, Joey Tianyi Zhou, Ran He, Anderson Rocha, Alex Kot

Keywords: Text-Centric, MLLM, Report generation, Forgery Analysis, AIGC

Abstract: The rapid evolution of AIGC has intensified threats from text-centric image manipulation. Current forensic research and challenges predominantly prioritize facial deepfakes, leaving the complex domain of text-centric forgery significantly under-explored. Moreover, existing methodologies mainly focus on binary classification or coarse localization, overlooking the critical need for fine-grained interpretability and sophisticated text editing generation. To bridge these gaps in both research scenarios and methodologies, we introduce GenText-Forensics, the first AI security challenge dedicated to text-centric multimedia forensics. Supported by RealText-V2, a large-scale multilingual dataset spanning diverse real-world scenarios, the challenge establishes a unified adversarial framework. Track 1 (Defense): Forgery Analysis Report Generation requires generating comprehensive forgery analysis reports that integrate detection, spatial grounding, and natural language explanation. Track 2 (Attack): AIGC Text-Image Editing focuses on adversarial AIGC text editing to expose detection vulnerabilities. This initiative aims to advance the development of Multimedia Security against evolving text-centric forgeries.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8