Explainable Deepfake Detection

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. Recent studies have attempted to enhance the interpretability of classification results by providing spatial manipulation masks or temporal forgery segments. However, due to the limitations of forgery datasets, the practical effectiveness of these methods remains suboptimal. The primary reason lies in the fact that most existing deepfake datasets contain only binary labels, with limited variety in forgery scenarios, insufficient diversity in deepfake types, and relatively small data scales, 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 1.4M+ forged samples and encompassing 80 distinct deepfake methods. The DDL design incorporates four key innovations: (1) Comprehensive Deepfake Methods (covering 7 different generation architectures and a total of 80 methods), (2) Varied Manipulation Modes (incorporating 7 classic and 3 novel forgery modes), (3) Diverse Forgery Scenarios and Modalities (including 3 scenarios and 3 modalities), and (4) Fine-grained Forgery Annotations (providing 1.18M+ precise spatial masks and 0.23M+ precise temporal segments). 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.

Deepfake detection remains a critical challenge in the era of advanced generative models, particularly as synthetic media becomes more sophisticated. In this study, we explore the potential of state of the art multi-modal (reasoning) large language models (LLMs) for deepfake image detection such as (OpenAI O1/4o, Gemini thinking Flash 2, Deepseek Janus, Grok 3, llama 3.2, Qwen 2/2.5 VL, Mistral Pixtral, Claude 3.5/3.7 sonnet) . We benchmark 12 latest multi-modal LLMs against traditional deepfake detection methods across multiple datasets, including recently published real-world deepfake imagery. To enhance performance, we employ prompt tuning and conduct an in-depth analysis of the models' reasoning pathways to identify key contributing factors in their decision-making process. Our findings indicate that best multi-modal LLMs achieve competitive performance with promising generalization ability with zero shot, even surpass traditional deepfake detection pipelines in out-of-distribution datasets while the rest of the LLM families performs extremely disappointing with some worse than random guess. Furthermore, we found newer model version and reasoning capabilities does not contribute to performance in such niche tasks of deepfake detection while model size do help in some cases. This study highlights the potential of integrating multi-modal reasoning in future deepfake detection frameworks and provides insights into model interpretability for robustness in real-world scenarios.

Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both. We propose FORGE, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM's native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, FORGE produces region-referential explanations answering fine-grained attribute queries ("Does the eyes/nose/mouth look real or fake?") and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness.