Abstract
Recent advances in AIGC have exacerbated the misuse of malicious deepfake content, making the development of reliable deepfake detection methods an essential means to address this challenge. Although existing deepfake detection models demonstrate outstanding performance in detection metrics, most methods only provide simple binary classification results, lacking interpretability. In critical domains such as law, interpretability is crucial for enhancing the credibility and authority of decisions. Recent studies attempt to improve the interpretability of classification results by providing spatial manipulation masks or temporal forgery segments. However, the practical effectiveness of these methods remains suboptimal due to limitations of the forgery data. Most current deepfake datasets predominantly offer binary labels, only a few datasets with localization annotations. However, they suffer from restricted forgery scenarios, limited diversity in deepfake types, and insufficient data scale, making them inadequate for complex real-world scenarios. To address this predicament, we construct a novel large-scale deepfake detection and localization (DDL) dataset containing over 1.8M forged samples and encompassing up to 80 distinct deepfake methods. The DDL design incorporates four key innovations: (1) Diverse Forgery Scenarios, (2) Comprehensive Deepfake Methods, (3) Varied Manipulation Modes, and (4) Fine-grained Forgery Annotations. Through these improvements, our DDL not only provides a more challenging benchmark for complex real-world forgeries, but also offers crucial support for building next-generation deepfake detection, localization, and interpretability methods.
Introduction: The Interpretability Gap
As generative AI (AIGC) explodes, deepfakes are becoming more sophisticated and harder to catch. Most existing detection models act as "black boxes"—they might identify a video as fake, but they cannot explain why. This lack of interpretability is a major hurdle for high-stakes applications like legal proceedings or news verification.
Current deepfake research faces four major gaps:
- Lack of Interpretability: Most datasets only provide binary labels (0 or 1), offering no insight into manipulated regions.
- Limited Diversity: Benchmarks often focus on older GAN-based methods, failing to keep up with newer architectures like Diffusion models, NeRFs, and VAEs.
- Small Scale: Many datasets are too small to train robust, real-world models.
- Narrow Scenarios: Most focus on single-face images, ignoring multi-face videos or complex audio-visual (AV) forgeries.
Methodology: A Human-in-the-Loop Pipeline
The researchers developed a human-in-the-loop generation pipeline to create the DDL dataset, ensuring it functions as a systematically engineered benchmark rather than just a collection of fakes.
- Unified Generation Pipeline: Driven by LLMs and human experts, the pipeline covers four scenarios: Single-face, Multi-face, Audio-Visual Full Synthesis, and Audio-Visual Local Manipulation.
- Novel Forgery Modes: The dataset introduces several new types of manipulation, including Hybrid Facial Forgery (HFF) (combining multiple face-swapping or reenactment techniques), Audio-Visual Asynchronous Manipulation (AVAM), and Audio-Visual Full Synthesis (AVFS).
- Automated Annotation: Unlike post-hoc labeling, DDL generates spatial masks and temporal segments during the synthesis process, ensuring pixel-perfect ground truth for localization tasks.
- Real-world Perturbations: To simulate the "messiness" of the internet, 27 types of perturbations (compression, noise, weather effects) were applied to the test set.
The DDL Dataset
The DDL Dataset is split into two main subsets: DDL-I (Image), focused on spatial localization, and DDL-AV (Audio-Visual), focused on temporal and cross-modal forgeries.
By the numbers: * 1.8M+ forged samples. * 80 distinct deepfake methods (including 34 GANs, 24 Diffusion models, VAEs, NeRFs, and commercial software). * 1.18M+ precise spatial masks and 0.23M+ precise temporal segments.
The dataset draws from 14 source datasets but intentionally limits the overlap between training and testing sources to only 21.05%, forcing models to evaluate Out-of-Distribution (OOD) performance and generalize to unseen faces.
Key Results and Insights
- Localization is the Key to Trust: By providing masks and segments, DDL allows detection models to transition from "guessing" to "reasoning," which is essential for forensic applications.
- The "Cat and Mouse" Game: By including 80 different methods, DDL highlights that a model trained only on GANs will likely fail against Diffusion-based forgeries. Diversity in the training set is non-negotiable for modern detectors.
- Real-World Validation: The authors integrated DDL into Ant Group’s AIGC detection platform. Online A/B testing showed a 95% detection accuracy across 80 different attack types in diverse international settings.
- Benchmark Superiority: DDL is currently the only dataset that covers Binary Classification (Cla), Spatial Localization (SL), and Temporal Localization (TL) across Image, Video, and Audio modalities simultaneously.
Conclusion and Future Work
DDL represents a significant step forward in moving deepfake detection beyond simple binary labels. While the current version provides the "where" and "when" of manipulation, it does not yet provide textual reasoning (e.g., explaining lighting inconsistencies). Future work will focus on integrating DDL with Vision-Language Models (VLMs) to generate natural language explanations, further bridging the gap between automated detection and human-understandable forensics.
Summary updated