DFF: InstructBLIP-based Explainable DeepFake Detection
📖 Model Description
This is the core DFF (DeepFake Detection and Forensic Explanation Framework) model as described in the ACL 2026 paper: "Generating Attribution Reports for Manipulated Facial Images: A Dataset and Baseline".
DFF is built upon the InstructBLIP (Flan-T5 XL) architecture. By integrating the Face-ViT auxiliary classifier, it achieves state-of-the-art performance in both forgery localization (mask generation) and forensic explanation (captioning).
🌟 Key Capabilities
- Forgery Localization: Generates high-resolution binary masks highlighting manipulated facial regions.
- Natural Language Explanation: Produces detailed text describing why a specific image is considered a forgery (e.g., "The texture around the eyes is unnatural due to GAN-based blending").
🛠️ Model Details
- Base LLM: Flan-T5 XL.
- Visual Encoder: EVA-ViT-G.
- Auxiliary Module: Face-ViT (Multi-label perception).
- Task: Explainable Detection & Multi-modal Attribution Reporting.
🚀 Links
- Official Code: Generating-Attribution-Reports
- Auxiliary Classifier: LianJC/Face-ViT-MultiLabel
- Dataset (MMTT): LianJC/MMTT-Dataset
📜 Citation
@inproceedings{lian2026generating,
title={Generating Attribution Reports for Manipulated Facial Images: A Dataset and Baseline},
author={Lian, Jingchun and others},
booktitle={Proceedings of ACL},
year={2026},
note={To appear}
}