SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection
SafeGuard employs multi-agent perception and reasoning, achieving 18.7% accuracy improvement in social-risk video detection on SafeVid dataset.
Key Findings
Methodology
SafeGuard integrates a hierarchical perceptual solver and a self-reflective verifier. The perceptual solver employs coarse-to-fine strategies, extracting fine-grained forensic cues such as optical flow, depth, appearance, and pixel anomalies using algorithms like Grounding DINO, SAM 2, and LVLM. The verifier cross-validates semantic hypotheses with physical plausibility, forming an interpretable evidence chain. This multi-modal, multi-step process enables comprehensive detection of AI-generated social-risk videos, combining visual forensic cues with high-level semantic reasoning.
Key Results
- On SafeVid, containing 20,000 videos across 10 social risk categories, SafeGuard achieved an accuracy of 85.3%, representing an 18.7% increase over previous methods. It outperformed benchmarks like LOKI, GenVideo, and DVF, especially in complex social scenarios. Ablation studies confirmed that multi-modal feature extraction and the closed-loop verification significantly contributed to performance gains.
- The system's interpretability stems from explicit evidence chains linking forensic cues and semantic hypotheses, allowing traceability of detection decisions. It demonstrated robustness across diverse high-risk scenarios, including violence, misinformation, and ethical violations.
- The framework's generalization was validated through cross-dataset tests, showing consistent improvements and reliable detection in unseen social contexts, highlighting its potential for real-world deployment.
Significance
This work addresses the critical gap in social-risk video detection by unifying perceptual forensic analysis with semantic reasoning. It advances the state-of-the-art by providing a transparent, interpretable, and robust framework capable of handling complex, high-stakes scenarios. The introduction of SafeVid as a benchmark promotes research focused on societal safety, encouraging development of models that can understand and verify social behaviors and physical plausibility simultaneously. The approach has broad implications for public safety, misinformation control, and content moderation, offering a scalable solution for detecting sophisticated AI-generated malicious videos.
Technical Contribution
The primary innovation lies in the multi-agent, hierarchical design that combines low-level forensic cue extraction with high-level semantic reasoning within a closed-loop framework. The use of multimodal tools for physical anomaly detection, coupled with a structured hypothesis-verification cycle, enhances detection reliability and interpretability. The system's modular architecture allows leveraging pre-trained models like Grounding DINO, SAM 2, and LVLM without extensive fine-tuning, making it adaptable and efficient. The construction of SafeVid dataset, covering diverse social risk categories, further supports the development of targeted detection algorithms. This integrated approach pushes the boundary of AI forensics by enabling joint physical and semantic analysis, a step toward more trustworthy AI systems.
Novelty
This is the first work to systematically integrate multi-modal forensic cues with semantic reasoning via a multi-agent, hierarchical framework for social-risk video detection. Unlike prior methods focusing solely on artifacts or semantics, SafeGuard bridges the perception-reasoning gap, providing a transparent evidence chain and iterative verification. Its novel closed-loop verification mechanism ensures high reliability and interpretability, especially critical in high-stakes social contexts. The creation of SafeVid dataset, tailored for social safety, further distinguishes this work from existing benchmarks.
Limitations
- The system's performance may decline with highly sophisticated or multi-modal adversarial manipulations that evade current forensic cues. Its reliance on pre-trained models could limit adaptability to unseen scenarios.
- Computational complexity remains high, posing challenges for real-time deployment in large-scale surveillance or streaming applications.
- Despite extensive coverage, SafeVid cannot encompass all social risk types, and the model's robustness needs further validation in real-world, noisy environments.
Future Work
Future research will focus on enhancing real-time inference capabilities, integrating adaptive learning for unseen manipulation types, and expanding the SafeVid dataset to include more nuanced social scenarios. Exploring reinforcement learning for multi-agent collaboration and developing lightweight models for deployment are also promising directions. Additionally, incorporating user feedback and explainability modules will further improve trustworthiness and practical usability in social safety applications.
AI Executive Summary
The rapid advancement of video synthesis technologies has brought about unprecedented challenges in detecting AI-generated content, especially in high-stakes social scenarios. Traditional detection methods, primarily based on perceptual artifacts, struggle to cope with fully coherent and highly realistic forgeries. Recognizing this gap, Wenlin Wu and colleagues introduced SafeGuard, a novel multi-agent framework that synergistically combines fine-grained forensic evidence extraction with high-level semantic reasoning.
SafeGuard employs a hierarchical perceptual solver that leverages advanced algorithms like Grounding DINO, SAM 2, and LVLM to localize and analyze suspicious regions within videos. It extracts physical and visual anomalies such as optical flow inconsistencies, depth distortions, and boundary artifacts. Complementing this, the self-reflective verifier cross-validates these forensic cues against semantic hypotheses about the scene, ensuring logical consistency and physical plausibility. This iterative process forms an interpretable evidence chain, enabling transparent decision-making.
The authors also curated SafeVid, a comprehensive dataset with 20,000 videos across 10 social risk categories, designed explicitly for high-risk scenario evaluation. Experiments demonstrate that SafeGuard outperforms existing methods by 18.7% accuracy on SafeVid and shows superior generalization across benchmarks like LOKI, GenVideo, and DVF. Its ability to jointly analyze visual artifacts and semantic context makes it particularly effective in detecting complex, socially sensitive forgeries.
This work significantly advances AI forensics, bridging the perception-reasoning gap, and providing a scalable, interpretable solution for social safety monitoring. Future directions include optimizing for real-time deployment, expanding dataset diversity, and enhancing robustness against sophisticated manipulations, promising a safer digital environment in the age of hyper-realistic synthetic videos.
Deep Dive
Plain Language Accessible to non-experts
想象你在一家工厂工作，工厂里有很多不同的工人，每个人都负责不同的任务。有的工人专门检查产品的细节，比如颜色、形状和标签，确保没有瑕疵；有的工人则负责检查整个生产流程是否合理，是否符合设计图纸。现在，如果有人偷偷在产品上做了手脚，比如换了标签或篡改了部分内容，单靠检查细节可能会发现一些微小的异常，但很难判断这些异常是否意味着产品被篡改。工厂还需要一个能理解整体流程和设计意图的管理者，来判断这些细节是否合理。SafeGuard就像这样一个工厂管理系统，它既能细致检查每个产品的细节，又能理解整个生产流程的合理性，确保每件出厂的产品都是真实可靠的。这种结合细节和整体理解的方法，让工厂的检测变得更全面、更可信。
ELI14 Explained like you're 14
想象你在学校里，有个老师要检查学生的作业是不是自己写的。只看字迹或内容可能会发现一些作弊的迹象，但有时候这些细节不足以判断。老师还需要考虑整个作业的主题和逻辑，看它是否合理、符合题意。SafeGuard就像这个老师，它用两种方式帮忙：一种是仔细检查作业中的细节，比如字迹、格式，找出可能的伪造痕迹；另一种是理解作业的整体内容和逻辑，看是否符合题目的要求。它们合作，确保作业既没有作弊，也符合逻辑。这样，老师就能更准确、更公平地判断学生的作业是否真实。这个系统就像一个聪明的老师，既细心又懂得大局。
Abstract
As video generation paradigms evolve from localized manipulation to full-scene synthesis, AI-generated video detection becomes increasingly challenging, as forgeries exhibit coherent global structure and high perceptual realism. However, existing benchmarks are biased toward perceptual fidelity and primarily evaluate detectors based on perceptual artifacts, providing limited coverage of scenarios that require reasoning about violations of physical laws, structural coherence, or social logic. This dataset bias shapes current approaches and results in a Perception-Reasoning Gap: artifact-centric models capture low-level statistical irregularities yet lack semantic inference, whereas vision-language models perform semantic reasoning but remain insensitive to fine-grained forensic cues. To bridge this gap, we propose SafeGuard, a multi-agent framework that enables collaborative specialization between forensic perception and semantic reasoning. A hierarchical perceptual solver extracts fine-grained forensic evidence, while a self-reflective verifier enforces consistency between semantic inference and physical plausibility, forming an interpretable evidence chain. To support evaluation, we introduce SafeVid, a novel AI-generated video detection benchmark comprising 20K videos spanning 10 social risk categories, designed to evaluate physical plausibility, structural consistency, and the rationality of social behaviors. Extensive experiments demonstrate the generalization of SafeGuard, improving accuracy on SafeVid by +18.7% and consistently outperforming prior methods across four public benchmarks.