How it Works
Veritas introduces a pattern-aware reasoning framework designed to improve the generalizability of deepfake detection. The reward model can be customized, with options including UnifiedReward-qwen-7B, Qwen3-VL-8B, or UnifiedReward-2.0-qwen3vl-8B, depending on performance requirements.
The framework is built upon the HydraFake dataset, a rigorously curated collection of 52,000 images spanning real and AI-generated content. The dataset is structured hierarchically, with training data containing abundant samples but limited forgery types, while evaluation splits are divided into four distinct levels: in-domain testing (14,000 images), cross-model testing (11,000 images), cross-forgery testing (12,000 images), and cross-domain testing (15,000 images). This structure ensures a comprehensive assessment of a detector’s ability to generalize across different scenarios, particularly in cross-forgery and cross-domain settings where most existing detectors perform poorly.
The reasoning process in Veritas is designed to be transparent and human-aligned, providing interpretable decision pathways. The training pipeline supports customization, allowing users to fine-tune the Veritas-Cold-Start model on their own datasets or integrate P-GRPO with binary-labeled data to refine detection capabilities.
Implementation-wise, Veritas relies on Python and integrates with tools like ms-swift and DeepfakeBench for inference and evaluation. The repository provides scripts for deploying models, running inference on single images or datasets, and testing performance against the HydraFake benchmark.
The HydraFake dataset itself is a critical component, featuring real images sourced from eight datasets and fake images generated using advanced techniques, including classic datasets, high-quality public datasets, and self-constructed deepfake data. The dataset’s hierarchical evaluation protocol is designed to expose weaknesses in existing detectors, particularly in scenarios requiring cross-forgery or cross-domain generalization.
For practical deployment, Veritas provides detailed instructions for setting up the environment, including dependency installation via Conda and pip. The training data is structured into JSON files (e.g., sft36k.json, mipo3k.json, pgrpo8k.json), which organize samples for each training method. Users can customize their detectors by fine-tuning Veritas-Cold-Start on proprietary data or by integrating P-GRPO with their own labeled datasets. The repository also encourages further research into novel GRPO-style algorithms and collaborative frameworks for small vision models.
Inference is streamlined through scripts that support both ms-swift and vLLM backends. Testing on the HydraFake dataset involves preparing the data in a specific directory structure, with separate folders for training, validation, and testing subsets. Vision model testing is facilitated through DeepfakeBench, with configurable parameters for detector and dataset settings.
Users can replace the reward model with alternatives like UnifiedReward-2.0-qwen3vl-8B to improve performance, and the training pipeline accommodates custom datasets by adjusting JSON file paths and label formats. This flexibility ensures that Veritas can be adapted to emerging threats in AI-generated content, including video deepfakes, as demonstrated by extensions like VideoVeritas and Veritas++, which target video and general image detection, respectively.
Use Cases
Veritas is primarily targeted at applications requiring generalizable deepfake detection, particularly in scenarios where AI-generated content must be authenticated with high confidence. The HydraFake dataset’s cross-domain and cross-forgery evaluation splits ensure that detectors trained with Veritas can handle real-world variability, reducing false positives and negatives in diverse environments.
One key use case is dataset auditing, where organizations need to verify the integrity of image datasets used for training or deployment. Veritas can identify AI-generated images within large collections, ensuring that downstream models are not exposed to synthetic data that could degrade performance or introduce biases.
Another application is provenance verification, where the origin of an image must be traced to prevent misinformation or fraud.
The framework’s extensibility also supports custom detector development, allowing researchers or organizations to fine-tune Veritas for niche applications. For instance, a detector specialized in identifying deepfakes from a specific generator (e.g., MidJourney) can be created by leveraging Veritas-Cold-Start and P-GRPO with targeted training data. This adaptability is further demonstrated by extensions like VideoVeritas, which applies similar reasoning principles to video deepfake detection, expanding Veritas’s utility beyond static images.
The HydraFake dataset and its rigorous evaluation protocol provide a standardized way to compare detectors across multiple dimensions, including cross-model, cross-forgery, and cross-domain performance. Researchers can also use Veritas as a foundation for exploring new reasoning patterns or reward models to improve generalization.
Key Features & Capabilities
Veritas’s pattern-aware reasoning framework is its defining feature, distinguishing it from traditional deepfake detection methods that rely solely on feature extraction or classification.
The second stage, P-GRPO, refines these patterns using a reward model, which can be customized based on the user’s requirements. This modular approach allows for incremental improvements, as demonstrated by the integration of newer reward models like UnifiedReward-2.0-qwen3vl-8B.
The HydraFake dataset is a critical asset, offering a comprehensive benchmark for deepfake detection. Its 52,000 images are divided into four evaluation splits, each testing a different aspect of generalization: in-domain, cross-model, cross-forgery, and cross-domain. This structure exposes the strengths and weaknesses of detectors in practical scenarios, where real-world data often deviates from training distributions.
Users can deploy models using vLLM and run evaluations on the HydraFake dataset with minimal setup, thanks to automated scripts that handle data preparation, model loading, and metric calculation. The repository also provides compatibility with DeepfakeBench for vision model testing, ensuring broad applicability.
Users can fine-tune the Veritas-Cold-Start model on proprietary datasets to create specialized detectors, or they can integrate P-GRPO with their own labeled data to enhance performance. This extensibility is further demonstrated by the development of Veritas++ and VideoVeritas, which expand the original model’s capabilities to general image and video detection, respectively.
The model’s reasoning process is structured to provide clear, human-aligned decision pathways, which can be inspected to understand why a particular image was classified as real or fake.
Performance-wise, Veritas has been recognized for its generalization capabilities, particularly in cross-forgery and cross-domain scenarios where many existing detectors fail. The ICLR 2026 Oral presentation of the original paper underscores its significance in the research community, while extensions like VideoVeritas and Veritas++ indicate ongoing efforts to broaden its scope. The repository’s permissive Apache 2.0 license further facilitates adoption, allowing researchers and organizations to build upon the framework without legal restrictions.