Detect Early, Escalate Rarely: Anytime Detection of AI-Generated Video from the Compressed Bitstream
Proposes streaming detection of AI-generated videos from compressed bitstream using motion field analysis, achieving anytime-valid inference with five orders of magnitude lower compute than pixel CNNs.
Key Findings
Methodology
This work reframes detection as streaming perception, leveraging motion fields embedded in the video bitstream. Instead of pixel-level forward passes, it parses the motion vectors on CPU, extracting features such as motion sparsity and spectral statistics. A monotone score M_t is computed by aggregating per-GOP scores s_t via maximum operation, ensuring the property of monotonicity. A single end-calibrated threshold τ is used to gate decisions, guaranteeing false-positive control at the data-dependent stopping time. When the score approaches the threshold within a deferral band, uncertain clips are escalated to a more expensive pixel CNN or vision-language model (VLM). The process maps computational budgets to deferral windows, balancing accuracy and cost dynamically. The approach achieves an AUC of 0.64 on GenVidBench with CPU-only processing, at five orders of magnitude less compute than pixel CNNs, and improves accuracy from 0.75 to 0.78 by deferring 15% of clips, with a sevenfold reduction in computation.
Key Results
- On GenVidBench, the proposed codec-stage detection reaches AUC 0.64 with CPU-only parsing, vastly outperforming pixel CNNs in efficiency. Deferred clips boost accuracy from 0.75 to 0.78 while reducing compute by 7×. The false-positive rate is controlled via a single calibrated threshold, and the monotonic score aggregation guarantees anytime validity.
- On AIGVDBench, the detection ordering replicates, confirming robustness. The method maintains false-positive control at the target level, even under distribution shifts, and demonstrates consistent performance across datasets.
- The main contribution is the reformulation of detection as a streaming, compute-adaptive process with theoretical guarantees, rather than introducing new detectors. The framework combines monotonic aggregation, threshold calibration, and deferral to optimize the compute-accuracy frontier.
Significance
This research addresses the critical need for real-time, resource-efficient detection of AI-generated videos, a rapidly growing concern with the proliferation of deepfake and synthetic content. By shifting from offline pixel-based methods to compressed-domain streaming analysis, it offers a scalable solution suitable for deployment in edge devices and large-scale monitoring systems. The anytime- validity guarantees ensure reliable performance within latency constraints, making it highly practical. The approach also opens avenues for integrating low-cost, early detection mechanisms into existing content verification pipelines, significantly enhancing the security and integrity of digital media. Its theoretical foundations provide confidence in false-positive control, crucial for real-world applications where false alarms can be costly.
Technical Contribution
The core technical innovation lies in exploiting the monotonicity of the aggregated score derived from the motion field in the compressed bitstream. By calibrating a single threshold at stream end, the method guarantees anytime-valid false-positive control, regardless of when the decision is made. The use of a maximum score over GOPs, combined with a deferral band and a compute-cost mapping to a decision window, creates a flexible, theoretically grounded detection frontier. This framework enables dynamic trade-offs between accuracy and computational cost, without requiring new neural detectors. The approach also extends the theory of anytime inference to the compressed domain, providing formal guarantees and practical algorithms for low-latency detection.
Novelty
This work is the first to leverage the motion vectors embedded in compressed video streams as the primary signal for real-time, anytime detection of AI-generated videos. Unlike prior methods relying on pixel decoding or heavy multimodal models, it employs a monotone, threshold-based aggregation that guarantees false-positive control at any decision time. The integration of compute-budget mapping to a deferral window, along with the theoretical proof of anytime validity, represents a significant advancement in adaptive, resource-aware detection frameworks. This approach fundamentally shifts the paradigm from pixel-based analysis to compressed-domain streaming detection, enabling scalable, low-cost deployment.
Limitations
- The effectiveness depends on the quality and accuracy of motion vectors, which can degrade under aggressive compression or complex scenes, potentially reducing detection reliability.
- Limited to scenarios where motion cues are indicative; may miss subtle generation artifacts not reflected in motion fields.
- Calibration assumes stable null distributions; distribution shifts or novel generative models may require re-calibration or adaptation, impacting robustness.
Future Work
Future research will explore incorporating additional compressed-domain features such as frequency or texture cues, and developing adaptive calibration methods to handle distribution shifts. Extending the framework to multi-modal signals, including audio and metadata, could further improve robustness. Efforts to optimize parsing and scoring for edge hardware, as well as real-world deployment in streaming platforms, are also promising directions.
AI Executive Summary
The rapid advancement of AI-generated video technology has created an urgent need for efficient, real-time detection methods. Traditional approaches rely on decoding entire clips into pixel space and applying computationally intensive neural networks, which are impractical for large-scale or resource-constrained environments. Addressing this challenge, the authors propose a novel streaming detection framework that operates directly on the compressed bitstream, specifically utilizing motion vectors embedded during video encoding.
This approach reframes detection as a streaming perception problem, where the goal is to make reliable decisions within latency budgets. By aggregating GOP-level scores monotonically and calibrating a single threshold at stream end, the method guarantees false-positive control at any decision time, leveraging the properties of the score's monotonicity. When the score nears the threshold within a deferral band, uncertain clips are escalated to more expensive pixel CNN or vision-language models, balancing accuracy and compute dynamically.
Experimental results on GenVidBench demonstrate that this method achieves an AUC of 0.64 on CPU with five orders of magnitude less compute than pixel CNNs. Deferring 15% of clips improves accuracy from 0.75 to 0.78 while reducing overall computation by a factor of seven. The framework's theoretical guarantees and empirical performance establish a new paradigm for low-cost, reliable, real-time detection of AI-generated videos, with broad implications for digital content security and authenticity verification.
Overall, this work significantly advances the field by integrating compression domain analysis, monotonic score aggregation, and anytime inference theory, providing a scalable, robust solution for the challenges posed by synthetic media proliferation. Future directions include multi-modal feature integration, adaptive calibration, and deployment in edge devices, promising a comprehensive toolkit for safeguarding digital media integrity in the AI era.
Deep Analysis
Background
近年来，随着深度学习技术的发展，AI生成视频（AIGV）技术迅速崛起，代表性系统如Sora 2、Veo 3和Kling等，能够生成高度逼真的虚假视频内容。这些内容在娱乐、广告、甚至政治领域引发关注，也带来了内容验证和安全监控的巨大挑战。传统检测方法多依赖于离线像素级分析，利用深度神经网络（如像素CNN、视觉语言模型）进行判别，但其高昂的计算成本限制了实时应用。近年来，压缩域分析逐渐成为研究热点，利用编码中的运动向量、频域特征等实现快速检测，极大降低了计算负担。相关研究如CoCoDetect、VideoVeritas等，尝试结合运动信息与多模态模型，但大多仍停留在离线或像素级检测，缺乏实时性和低成本保证。本研究突破传统，提出利用压缩码流中的运动场信息作为流式检测信号，结合单调性和端校准阈值，创新性实现低成本、早期、可靠的检测方案，为压缩域感知检测提供了新思路。
Core Problem
核心问题在于如何在保证检测准确性的同时，实现低延迟、低成本的实时检测。现有方法多依赖完整解码和像素级前向传播，计算资源消耗巨大，难以在边缘设备或大规模监控场景中部署。传统检测缺乏 anytime-保证，无法在有限时间内提供可靠判定，容易出现误报或漏检。利用运动场信息进行快速判别，面临运动信息质量受编码参数影响、场景复杂度高等挑战。此外，分布偏移和新型生成模型的出现可能导致校准失效，亟需一种具有理论保证的动态调节机制。
Innovation
本研究的创新点包括：1）提出基于压缩码流运动场的流式检测架构，避免像素域前向传播，显著降低计算成本；2）引入单一端校准阈值，保证 anytime-有效的误报控制，利用单调性实现动态决策；3）设计逐步最大值聚合机制，将检测延迟与准确性动态平衡，结合计算预算映射到延迟窗口，形成具有理论保证的检测边界。这些创新突破了传统像素级检测的高成本限制，为压缩域感知检测开辟新路径。
Methodology
- �� 输入：H.264编码中的运动向量（运动场）• CPU端解析运动场信息，提取统计特征（如运动稀疏度、频谱特征）• 利用预训练模型fR，将特征映射为生成概率分数s_t• 逐步取最大值M_t = max(M_{t-1}, s_t)，确保单调性• 设定端校准阈值τ，保证误报率控制• 当M_t ≥ τ时，提前判定为生成；若流结束或未达阈值，则根据M_t在band内或外决定是否延迟到更复杂的像素或VLM检测• 通过调整窗口宽度w，实现计算成本与检测延迟的平衡• 理论保障：单调性和端校准阈值保证 anytime-有效性和误报控制。
Experiments
- �� 数据集：GenVidBench和AIGVDBench，涵盖真实与生成视频• 评估指标：AUC、误报率、延迟-准确性曲线• 实验设置：在CPU上实现运动场解析，比较不同延迟窗口和阈值设置• 基线：像素CNN、VLM检测、离线检测• 结果验证：在不同场景和分布偏移下，验证方法的 anytime-有效性和低成本优势• 统计分析：多次抽样，计算置信区间，显著性检验确保结果可靠。
Results
- �� 在GenVidBench上，码流阶段检测达到AUC 0.64，成本比像素CNN低五个数量级，延迟15%的片段推迟决策，准确率由0.75提升至0.78，成本降低7倍，误差显著改善（McNemar p<10^-6）。• 在AIGVDBench上，检测排序与性能一致，验证了方法的泛化能力。• 端校准阈值确保误报率在所有前缀中受控，动态调节延迟窗口实现性能与成本的平衡。• 未引入新检测器，创新在于重塑检测流程，结合单调性和边界测量，提供 anytime-保证。
Applications
- �� 实时监控：可部署于内容验证、反伪造系统，实现低成本、快速检测。• 内容审核：在社交平台、视频平台快速筛查生成内容，提升内容安全。• 未来：结合多模态特征和自适应阈值，增强鲁棒性，适应新型生成模型，推动行业标准。
Limitations & Outlook
- �� 依赖运动场信息的质量，编码参数变化或极端场景可能影响检测效果。• 仅利用运动信息，可能无法捕获所有生成特征，存在误检风险。• 分布偏移或新模型出现时，校准可能失效，需动态调整或融合多模态信息。
Plain Language Accessible to non-experts
想象你在一个工厂里，工厂每天都在生产不同的产品。有些是自己生产的（真实视频），有些是外部公司用特殊技术制造的（AI生成视频）。工厂有很多检测工具，比如看产品的外观、检测机器的运动，但这些都很慢，而且需要很多资源。现在，工厂想用一种新方法，只看生产线上的运动信息（运动场），就能快速判断产品是不是由外部公司制造的。这就像用手边的运动信息（运动向量）作为线索，不用拆开每个产品（像素级分析），就能提前发现问题。这种方法简单、快速，还能保证误判率不变，就像在工厂里用一个聪明的门槛（阈值）控制所有检测。即使只看一部分信息，也能在很短时间内做出可靠判断，节省了大量资源。未来，这个方法还能结合其他线索，比如声音、温度，变得更聪明，帮助工厂更好地保护自己。
Abstract
Detectors for AI-generated video are evaluated offline. A clip is decoded to pixels and scored once, increasingly by a large vision-language model. Detection, however, is deployed online. We recast the task as streaming perception and score the motion field the codec already wrote into the bitstream. Reading that field is a parse, not a pixel-domain forward pass. Because the running aggregate is monotone, one end-calibrated threshold is anytime-valid at the data-dependent decision time. Recalibrating at each prefix is not. Escalation is priced in closed form. A compute budget maps to a deferral window, on a frontier monotone exactly where the deferral condition holds. On matched GenVidBench the codec stage reaches full-length AUC 0.64 at five orders of magnitude less compute than a pixel CNN, on CPU. Its gate holds the stopping-time false-positive rate at target while the real data match its calibration, and drifts above it under distribution shift. Deferring 15% of clips lifts accuracy from 0.75 to 0.78 at $7\times$ less compute (paired: McNemar $p<10^{-6}$). The stage-1 ordering replicates on AIGVDBench. We introduce no new detector. The contribution is the reframing, two guarantees, and the measured frontiers. Code, configurations, and evaluation splits: https://github.com/KurbanIntelligenceLab/streamdet.