Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning

TL;DR

This work introduces HydraFake, a dataset that simulates real-world challenges with hierarchical generalization testing, and proposes Veritas, a multi-modal large language model (MLLM) based deepfake detector capable of delivering transparent and faithful detection outputs.

Abstract

Deepfake detection remains a formidable challenge due to the complex and evolving nature of fake content in real-world scenarios. However, existing academic benchmarks suffer from severe discrepancies from industrial practice, typically featuring homogeneous training sources and low-quality testing images, which hinder the practical deployments of current detectors. To mitigate this gap, we introduce HydraFake, a dataset that simulates real-world challenges with hierarchical generalization testing. Specifically, HydraFake involves diversified deepfake techniques and in-the-wild forgeries, along with rigorous training and evaluation protocol, covering unseen model architectures, emerging forgery techniques and novel data domains. Building on this resource, we propose Veritas, a multi-modal large language model (MLLM) based deepfake detector. Different from vanilla chain-of-thought (CoT), we introduce pattern-aware reasoning that involves critical reasoning patterns such as "planning" and "self-reflection" to emulate human forensic process. We further propose a two-stage training pipeline to seamlessly internalize such deepfake reasoning capacities into current MLLMs. Experiments on HydraFake dataset reveal that although previous detectors show great generalization on cross-model scenarios, they fall short on unseen forgeries and data domains. Our Veritas achieves significant gains across different OOD scenarios, and is capable of delivering transparent and faithful detection outputs.

Figures (43)

• Figure 1: Comparison of the detection outputs. InternVL3-78B zhu2025internvl3 gets incorrect answer. GPT-4o hurst2024gpt and our model trained without the proposed MiPO both fail to provide precise explanation. In contrast, our model gives transparent and faithful decision process.
• Figure 2: Overview of HydraFake dataset. (a) We carefully collect and reimplement advanced deepfake techniques to construct our HydraFake dataset. Real images are collected from $8$ datasets. Fake images are from classic datasets, high-quality public datasets and our self-constructed deepfake data. (b) We introduce a rigorous evaluation protocol. Training data contains abundant samples but limited forgery types. Evaluations are split into four levels. (c) Illustration of the subsets in different evaluation splits. (d) Performance of existing detectors. Most detectors generalize well on Cross-Model setting but perform poorly on Cross-Forgery and Cross-Domain scenarios.
• Figure 2: Effect of the proposed pattern-aware reasoning.
• Figure 3: Overview of two-stage training pipeline. (a) For Pattern-Guided Cold-Start, we first employ SFT to internalize thinking patterns. Then we introduce MiPO to facilitate human-aligned reasoning. The MiPO dataset consists of mixed non-preference data, encouraging model to perform precise and fine-grained reasoning. (b) For Pattern-Aware GRPO, we introduce pattern-aware reward to incentivize adaptive reasoning ability on pattern granularity, yielding our Veritas$\,$ model.
• Figure 4: Ablations on the training stages. "Avg" is directly averaged across four splits.