PaperlayerView exact revision on arXiv

Source-linked AI summary

DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark

Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, Huaxiong Li

arXiv:2405.19707v3cs.CV

TL;DR

AI-generated videos raise misinformation concerns, but detection research and datasets have been limited for generated videos beyond human faces. This paper introduces GenVideo and two evaluation tasks, alongside DeMamba, a plug-and-play module for spatial-temporal inconsistencies; the authors report improved generalizability and robustness on GenVideo. [The passage-grounded facets below provide the supporting claims.]

  • Problem

    Detection studies have addressed deepfake videos, but there is a lack of research on generated videos beyond human faces; large-scale datasets from advanced generators are also lacking.

    Source: §1 · p. 2, §2.2 · p. 4

  • Method

    The paper introduces the diverse, million-scale GenVideo dataset, cross-generator and degraded-video evaluation tasks, and DeMamba for modeling spatial-temporal inconsistencies.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    DeMamba is reported to enhance detectors’ generalizability and robustness on GenVideo.

    Source: §1 · p. 2, §7 · p. 12

  • Takeaways & Limitations

    The authors present GenVideo and DeMamba as contributions to developing AI-generated video detection technologies.

    Source: §7 · pp. 12–13

    The authors identify DeMamba’s suboptimal training efficiency, a common issue with Mamba models, as the paper’s main limitation.

    Source: §7 · p. 13

Abstract

from arXiv · show

Recently, video generation techniques have advanced rapidly. Given the popularity of video content on social media platforms, these models intensify concerns about the spread of fake information. Therefore, there is a growing demand for detectors capable of distinguishing between fake AI-generated videos and mitigating the potential harm caused by fake information. However, the lack of large-scale datasets from the most advanced video generators poses a barrier to the development of such detectors. To address this gap, we introduce the first AI-generated video detection dataset, GenVideo. It features the following characteristics: (1) a large volume of videos, including over one million AI-generated and real videos collected; (2) a rich diversity of generated content and methodologies, covering a broad spectrum of video categories and generation techniques. We conducted extensive studies of the dataset and proposed two evaluation methods tailored for real-world-like scenarios to assess the detectors' performance: the cross-generator video classification task assesses the generalizability of trained detectors on generators; the degraded video classification task evaluates the robustness of detectors to handle videos that have degraded in quality during dissemination. Moreover, we introduced a plug-and-play module, named Detail Mamba (DeMamba), designed to enhance the detectors by identifying AI-generated videos through the analysis of inconsistencies in temporal and spatial dimensions. Our extensive experiments demonstrate DeMamba's superior generalizability and robustness on GenVideo compared to existing detectors. We believe that the GenVideo dataset and the DeMamba module will significantly advance the field of AI-generated video detection. Our code and dataset will be aviliable at \url{https://github.com/chenhaoxing/DeMamba}.

1 Introduction

The paper addresses the need to detect AI-generated videos and introduces GenVideo, evaluation tasks, and DeMamba. It presents DeMamba as a plug-and-play module for modeling spatial-temporal inconsistencies and improving detector generalizability and robustness.

1 Introduction

  • Generated videos can be used to misinform the public, and an AI-generated Taylor Swift video spread widely on Twitter and harmed her reputation.The paper argues that this motivates technology for detecting fake videos and avoiding potential harm.Source: §1 · pp. 1–2
  • GenVideo contains over one million generated videos and over one million real videos, spanning varied scenes, content, and generation methods.The paper proposes cross-generator classification for unseen generators and degraded-video classification for videos affected by quality degradation.Source: §1 · p. 2
  • DeMamba models spatial-temporal inconsistencies and is reported to improve existing detectors’ generalizability and robustness on GenVideo.The module is designed as a plug-and-play addition to feature extractors.Source: §1 · p. 2
  • The authors identify a mismatch in existing detectors: image methods cannot model temporal inconsistencies, while video methods struggle to model local spatial inconsistencies.They argue that satisfactory generalization requires modeling spatial-temporal local details.Source: §1 · p. 2

2 Related works

Prior video-generation work spans text- and image-conditioned tasks and several model families, while detection research has largely focused on generated images or face-focused deepfakes. The paper identifies broader AI-generated video detection beyond human faces as an underexplored area.

2.1 Video generation methods

  • Video generation includes text-to-video and image-to-video tasks, with methods grouped into diffusion U-Nets, diffusion transformers, and other approaches.The related work describes examples across these families, including GANs and autoregressive models.Source: §2.1 · pp. 2–3
  • Generated videos can show local-appearance errors, frequency inconsistency, and temporal inconsistency.Figure 1 illustrates these spatial and temporal artifacts.Source: Figure 1 · p. 3

2.2 AI-Generated content detection

  • Prior detection work includes generated-image detection and deepfake-video detection, but the cited video studies focus on human faces rather than broader generated-video content.The paper positions detection of generated videos beyond human faces as an area needing research.Source: §2.2 · p. 4

3 GenVideo

GenVideo is designed to support evaluation of detector generalizability and robustness with a large, diverse collection and two real-world-oriented tasks. Its construction draws on existing datasets, generation pipelines, and scraped videos from varied models.

3.1 Overviews of GenVideo

  • GenVideo’s training and testing sets contain 2,294,594 and 19,588 clips, respectively, and the dataset evaluates cross-generator generalization and classification of degraded videos.Cross-generator evaluation tests unseen generators; degraded-video classification tests recognition under different types of degradation.Source: §3.1 · pp. 4–5

3.2 Organization of GenVideo

  • The dataset uses real videos from existing video datasets and generated videos from web scraping, in-house generation, and existing evaluation datasets.Its training set emphasizes diffusion and autoregressive methods, while the testing set includes MSR-VTT real videos and generated videos from EvalCrafter and web scraping.Source: §3.2 · p. 5

3.3 Video collection details of GenVideo

  • GenVideo’s collection aims to maximize diversity in content and generators, using text-to-video and image-to-video pipelines and scraped examples from current models.The generation pipeline uses expanded prompts across foreground categories and scenes, while the testing collection includes popular and non-mainstream generators.Source: §3.3 · pp. 5–6

4 DeMamba

DeMamba combines input-adaptive state-space modeling with spatially continuous scans of video-feature regions to model spatial-temporal inconsistencies. It classifies videos using both global and local features.

4.1 Preliminaries

  • S4 maps an input sequence x(t) to y(t) through hidden state h(t); Mamba discretizes its dynamics and derives S6 parameters from the input.A represents system evolution, B and C are projection matrices, and the timescale parameter Δ supports Zero-Order Hold discretization.Source: §4.1 · pp. 6–7

4.2 AI-generated video detection with DeMamba module

  • Vision encoders transform video frames into feature maps, which DeMamba divides into spatial zones for intra-group feature consistency modeling.The resulting local features are aggregated to determine whether the video is AI-generated.Source: §4.2 · p. 7
  • DeMamba scans each zone in a spatially continuous order, aligning nearby spatial tokens sequentially across frames to model spatial-temporal inconsistencies.For a four-position zone, its continuous order is (1,1) → (1,2) → (2,2) → (2,1), unlike the sweep-scan order (1,1) → (1,2) → (2,1) → (2,2).Source: §4.2 · p. 7
  • The classification head concatenates global features averaged before DeMamba with pooled local features, then uses an MLP trained with binary cross-entropy.The global feature averages temporal and spatial dimensions; the local feature pools temporal and spatial features after DeMamba.Source: §4.2 · p. 7

5 Experiments

The experiments evaluate detector generalization across generators and robustness to degraded videos, alongside ablations of DeMamba. DeMamba-XCLIP-FT performs better across the reported generalization and degradation evaluations, while ablations examine design choices and efficiency.

5.1 Implementation details

  • Training uses 1,213,511 real and 1,048,575 generated videos; the out-of-domain test set has 10,000 real and 8,588 generated videos.The training set includes videos from 10 baseline generative methods.Source: §5.1 · p. 7

5.2 Task1: cross generator generalization

  • DeMamba-XCLIP-FT achieves average recall/F1/AP of 0.9392/0.9020/0.9710, improving 11.26%/17.72%/12.02% over original XCLIP.These are results for the many-to-many generalization task.Source: Table 3 · p. 8, §5.2 · p. 10
  • DeMamba-XCLIP-FT achieves better generalization performance in all three one-to-many generalization tasks.The task trains on one baseline generator and tests on each subset and the average out-of-domain test performance.Source: Table 4 · p. 9, §5.2 · p. 10

5.3 Task2: degraded video classification

  • Under eight types of perturbations, DeMamba-XCLIP-FT achieves the best performance on degraded data.The perturbations include H.264 and JPEG compression, FLIP, cropping, watermarks, Gaussian noise, and color transform.Source: Table 5 · p. 10, §5.3 · p. 11

5.4 Ablation study

  • DeMamba improves generalization, fused features yield the best performance, and continuous scanning improves on traditional scanning.The ablations report the fused-feature result and continuous-scan comparison separately.Source: Table 6 · p. 10, §5.4 · p. 11, Table 8 · p. 11
  • Zone size 2 gives the best performance; smaller zones emphasize local details, but excessively small zones can lose spatial context.The authors emphasize selecting an appropriate zone size.Source: Table 7 · p. 11, §5.4 · p. 11
  • Detection performance improves in association with dataset scale across experiments varying training-sample counts and generator counts.The comparisons use the full test set and DeMamba-XCLIP-FT.Source: §5.4 · p. 12
  • Adding DeMamba to XCLIP does not significantly affect inference time; training is slower, but posttraining inference is efficient and computational times remain competitive with video models.The reported comparisons include outperforming F3Net, STIL, and TALL in computational times.Source: Table 9 · p. 12, §5.4 · p. 12

6 Broader impacts

The authors frame machine-learning detection of generated videos as a means of protecting digital content and limiting misinformation harms, while noting the potential misuse of detection tools.

6 Broader impacts

  • The authors present generated-video detection as a way to protect digital content and prevent misinformation, while acknowledging that detection tools could be misused.They advocate ethical technology use and research into media-authenticity tools, with attention to personal privacy.Source: §6 · p. 12

7 Conclusion and limitation

The paper presents GenVideo and DeMamba as a dataset and detector for AI-generated video, evaluates detection in cross-generator and degraded-video tasks, and reports strong generalization and robustness. It identifies DeMamba’s suboptimal training efficiency as its main limitation.

7 Conclusion and limitation

  • DeMamba demonstrates strong generalization and robustness across multiple tasks.The conclusion also describes it as distinguishing AI-generated videos through spatial-temporal inconsistencies.Source: §7 · p. 12
  • GenVideo supports cross-generator and degraded-video classification to evaluate detectors in real-world-like scenarios.The dataset is described as large-scale and diverse in generated content and methods.Source: §7 · p. 12
  • The authors hope the work will inspire detection technologies for authentic and reliable AI-generated-content applications.Source: §7 · p. 13
  • DeMamba’s training efficiency is suboptimal, a limitation the paper describes as common to Mamba models.The authors encourage lightweight and generalized detection models.Source: §7 · p. 13

A Model details

  • DeMamba adds only a small number of parameters to XCLIP-B while achieving significant performance improvements.Table 10 reports model details and performance on the many-to-many generalization task.Source: A Model details · p. 20, Table 10 · p. 20

B.1 Implementation details

The implementation samples video frames according to video length, then selects consecutive frames, resizes them, and applies random training augmentations.

B.1 Implementation details

  • Videos longer than 3 seconds are sampled at 2 frames per second, with frame extraction branching at a 3-second threshold.Source: §B.1 · p. 20
  • After frame sampling, training and testing randomly select 8 or 16 consecutive frames and resize each frame to 224 × 224.Source: §B.1 · p. 20
  • Training uses random HorizontalFlip, ImageCompression, GaussNoise, GaussianBlur, and Grayscale augmentations.The stated purpose is to enhance model generalizability.Source: §B.1 · p. 20

B.2 Details in degraded video classification task

For degraded-video classification, the paper applies specified transformations to out-of-distribution videos and tests models trained on many-to-many tasks.

B.2 Details in degraded video classification task

  • The degradation setup tests many-to-many-trained models on transformed Dv−ood videos.Source: §B.2 · p. 21
  • The study compresses video with H.264 (also known as Advanced Video Coding, or AVC) at CRF 28, and describes JPEG as a widely used image-compression standard.Source: §B.2 · p. 21
  • Flips are selected between horizontal and vertical with equal probability, and crops retain 71%–93% of the original scale.Source: §B.2 · p. 21
  • The setup randomly adds text or image watermarks and applies Gaussian blur with σ = 7.Source: §B.2 · p. 21
  • Color degradation randomly selects brightness, contrast, saturation, or hue, with the parameter set to 0.5.Source: §B.2 · p. 21

C Visualization of dataset

The dataset visualization section presents generated-video samples from a range of models and notes that the samples show diverse content.

C Visualization of dataset

Related papers

Loading 2405.19707v3…