PATE-Forensics: Perception-as-Tool for Explainable Deepfake Forensics with General-Purpose MLLMs

Modular approach improves deepfake detection: PATE-Forensics separates the task into specialized components—a forensic "perception tool" handles detection and localization, while a general-purpose AI model generates explanations. This architectural separation achieved a score of 0.89 on the DDL-X Track 3 benchmark, outperforming the second-place team by 0.19 points, suggesting that specialized tools may be more effective than single models trying to do everything at once.

Explainability doesn't require task-specific training: The system uses a general-purpose multimodal language model (MLLM) to explain deepfake detection results without requiring specialized fine-tuning on forensic tasks. This approach makes the system more adaptable and potentially easier to maintain, as improvements to general AI models could automatically benefit deepfake forensics.

Multiple levels of evidence strengthen detection: The forensic perception tool integrates detection signals at three different scales—global image-level, patch-level (small regions), and segment-level (medium regions)—to identify manipulated content. This multi-granularity approach provides more robust evidence than relying on a single analysis method.

Structured context improves AI explanations: By providing the general-purpose language model with organized forensic outputs (detection maps and localization masks) rather than raw images, the system guides more accurate and relevant explanations of why content is flagged as deepfake, supporting more trustworthy AI decision-making.