Research map / Vision-language and multimodal / Multimodal LLMs
Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the HydraFake deepfake detection dataset with hierarchical OOD protocol and Veritas, an InternVL3-8B-based MLLM trained with pattern-guided cold-start (SFT + MiPO) and pattern-aware GRPO to produce transparent reasoning for deepfake detection.
Problem
Deepfake detectors trained on narrow sources and tested on homogeneous benchmarks generalize poorly to unseen forgery types and data domains, and existing MLLM-based detectors rely on post-hoc explanations rather than reasoning that drives the decision.
Approach
A pattern-aware reasoning format (fast judgement, planning, reasoning, self-reflection, conclusion) internalized into an MLLM via SFT, mixed-negative preference optimization (MiPO), and GRPO with a reward that favors planning/self-reflection when the answer is correct and penalizes them when wrong.
Key results
- Veritas achieves 90.7% average accuracy on HydraFake versus 84.7% for the best prior detector (Co-SPY), and reports 6.0% averaged gain over the previous best.
- Veritas surpasses Gemini-2.5-Pro by 11.8% on average and its base InternVL3-8B by 32.4%.
- Restricted-scope Veritas-mini reaches 85.8% average, outperforming MLLM-based detectors such as FakeVLM (77.3%) and SIDA-7B (76.3%).
Surprising findings
- Existing small-vision detectors generalize well on cross-model splits but drop below 85% on cross-forgery and cross-domain splits.
- Pure RL without cold-start underperforms the cold-started model, attributed to low-quality explorations causing unstable training.
Limitations
- Real-image quality in low-resolution datasets limits reliable self-reflection on authentic images, as noted in annotation pipeline.
- Cross-domain and cross-forgery results rely on reproduced deepfakes and a constructed dataset; some baselines were reproduced via DeepfakeBench with own training settings.
Experimental setup
Datasets: HydraFake FF++ CelebA LFW FFHQ CelebAHQ DF40 FFIW WILD TalkingHeadBench
Benchmarks: HydraFake
Metrics: accuracy precision recall
Baselines: F3Net UniFD IID FreqNet ProDet NPR AIDE Co-SPY D^3 Effort Qwen2.5-VL-7B InternVL3-8B
Builds on: GRPO DPO DeepSeek-R1 DeepfakeBench
Closest papers by meaning
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.91)
- Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.90)
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.90)
- Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.90)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.89)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.