A Two-Stage Agentic Framework for Explainable Deepfake Detection in the 2026 ACM Multimedia TRIDENT Image Track
Abstract
Recent advances in generative models have made facial image manipulation increasingly realistic, requiring deepfake detection systems to move beyond binary decisions and provide reliable artifact perception and faithful explanations. To address this challenge, we propose a two-stage agentic framework for the TRIDENT Image Track. In the first stage, we fine-tune a DINOv3-based facial foundation encoder to produce a reliable authenticity prior for real/fake detection. Conditioned on this prior, the second stage employs a fine-tuned VLM agent to perform artifact-aware reasoning for True-False Questions, Multiple-Choice Questions, and Open-Ended Questions. To improve explanation quality and reduce hallucinated artifact claims, we construct high-quality SFT data through self-generation and evaluator-guided self-reflection, followed by GRPO-based post-training. During inference, a hierarchical agentic workflow delegates different question types to task-specific specialist agents, while a final aggregator enforces output consistency, format compliance, and conservative artifact prediction. This design decouples authenticity estimation from language-based forensic reasoning while retaining consistency between detection and explanation. Our final system achieves a TCS score of 86.18 on the TRIDENT Image Track leaderboard, ranking first among all participating teams and demonstrating the effectiveness of combining specialized deepfake detection with agentic VLM reasoning for explainable deepfake understanding.