PaperlayerView exact revision on arXiv

Source-linked AI summary

Detecting AI-Generated Video via Frame Consistency

Long Ma, Zhiyuan Yan, Qinglang Guo, Yong Liao, Haiyang Yu, Pengyuan Zhou

arXiv:2402.02085v8cs.CVcs.AI

TL;DR

AI-generated video detection remains an overlooked challenge amid concerns about validating video authenticity. The paper introduces the GVF dataset and DeCoF, a detector focused on temporal artifacts, and reports generalization to unseen generators, including commercial models.

  • Problem

    AI-generated video detection is an overlooked challenge in validating video authenticity as text-to-video models emerge.

    Source: §VI · p. 5, J. Limitation · p. 13

  • Method

    The paper creates the GVF dataset and proposes DeCoF, which separates temporal from spatial artifacts to detect inter-frame anomalies.

    Source: §I · pp. 1–2

  • Results

    DeCoF generalizes to unseen generation models, including Sora, Veo, and Kling, and shows robustness across open-source and commercial closed-source models.

    Source: §I · p. 2, §VI · p. 5

  • Takeaways & Limitations

    GVF provides a public benchmark for community research on AI-generated video detection.

    Source: §VI · p. 5

    The method targets generated videos and may degrade or be inapplicable for other video forgeries, including Deepfakes and malicious editing.

    Source: J. Limitation · p. 14

Abstract

from arXiv · show

The escalating quality of video generated by advanced video generation methods results in new security challenges, while there have been few relevant research efforts: 1) There is no open-source dataset for generated video detection, 2) No generated video detection method has been proposed so far. To this end, we propose an open-source dataset and a detection method for generated video for the first time. First, we propose a scalable dataset consisting of 964 prompts, covering various forgery targets, scenes, behaviors, and actions, as well as various generation models with different architectures and generation methods, including the most popular commercial models like OpenAI's Sora and Google's Veo. Second, we found via probing experiments that spatial artifact-based detectors lack generalizability. Hence, we propose a simple yet effective \textbf{de}tection model based on \textbf{f}rame \textbf{co}nsistency (\textbf{DeCoF}), which focuses on temporal artifacts by eliminating the impact of spatial artifacts during feature learning. Extensive experiments demonstrate the efficacy of DeCoF in detecting videos generated by unseen video generation models and confirm its powerful generalizability across several commercially proprietary models.

I. INTRODUCTION

The introduction frames AI-generated video detection as an urgent response to realistic synthetic videos and their security risks. It presents the GVF dataset and DeCoF, a detector focused on temporal artifacts, and reports generalization to unseen generators.

I. INTRODUCTION

  • Realistic generated videos raise security concerns including privacy violations and erosion of trust on social media.The introduction argues that these risks make reliable detection necessary.Source: §I · p. 1
  • Naively trained spatiotemporal networks can rely on easy spatial artifacts and generalize poorly to unseen generators.The introduction contrasts these spatial shortcuts with subtler, potentially more generalizable temporal artifacts.Source: §I · p. 1
  • DeCoF targets temporal forgery artifacts by decoupling them from spatial artifacts during feature learning.The method maps frames to a feature space where inter-feature distance is inversely correlated with image similarity, focusing on anomalies between features.Source: §I · pp. 1–2
  • Experiments report that DeCoF generalizes to unseen generation models, including commercial models Sora, Veo, and Kling.The paper also validates the approach on the latest commercial proprietary models.Source: §I · p. 2
  • The paper introduces GVF as a comprehensive, scalable dataset for benchmarking AI-generated video detectors.Its real and generated videos use aligned prompts, and it includes videos from nine text-to-video generators with different methods.Source: §I · pp. 1–2

II. RELATED WORK

The related-work section reviews text-to-video diffusion models and prior generated-content detection, noting that AI-generated video detection remained unexplored. The dataset section describes GVF's prompt-based content coverage and generator diversity.

A. Generation Models

  • Text-to-video diffusion models include latent-based approaches such as Text2Video-zero and ModelScopeT2V, and pixel-based approaches such as PYoCo and Make-A-Video.Show1 combines pixel-based generation with latent-based upscaling.Source: A. Generation Models · p. 2

B. Generated Content Detection

  • Prior generated-image detectors and Deepfake-video studies had not addressed AI-generated video detection, which this paper targets as an exploratory direction.The cited image-detection work includes classifiers, artifact-based methods, diffusion reconstruction, and vision-to-language models.Source: B. Generated Content Detection · p. 2

III. DATASET CONSTRUCTION

  • GVF comprises 964 triples and samples prompts and real videos across spatial and temporal content and attributes.It considers people, animals, plants, food, vehicles, buildings, and artifacts, alongside actions, motion, fluid motion, and light change; eight open-source and commercial generators supply generated videos.Source: §III · p. 2

IV. METHODOLOGY

Probe experiments show that spatiotemporal detectors rely heavily on spatial artifacts, while image-based detectors capture artifacts specific to seen generation models rather than generalizable cues. The evidence motivates the difficulty of detecting unseen generators through spatial artifacts alone.

A. Probe experiments of STNN models

  • STNNs overly rely on spatial artifacts, reducing AI-generated video detection to a 2D problem.The probing experiments test temporal discontinuity using frame-scrambled real videos and spatial artifacts using repeated frames from fake videos.Source: A. Probe experiments of STNN models · p. 3, Table I · p. 3

B. Capturing Universal Spatial Artifacts is Difficult

  • Image-based detectors capture spatial artifacts from specific generation models but lack generalizability to unseen models.The feature-space analysis distinguishes real videos from videos generated by a seen model, but not fake videos from unseen models.Source: B. Capturing Universal Spatial… · p. 3, Figure 4 · p. 3
  • Spatial artifacts differ across generation models, making universal spatial-artifact detection difficult.The analysis notes parallel slashes in Text2Video-zero videos and bright spots in Show-1 videos.Source: B. Capturing Universal Spatial… · p. 3

C. Detection based on Temporal Artifacts

The method targets temporal artifacts associated with video continuity, while reducing the influence of spatial artifacts on feature learning. DeCoF uses frame features and a verification module to classify video authenticity based on frame consistency.

C. Detection based on Temporal Artifacts

V. EXPERIMENTS

The experiments use a training sub-dataset and remaining sub-datasets for generalizability evaluation. They report video-level accuracy and AUC, with a 0.5 threshold for accuracy.

A. Experimental Setup

  • The detector is trained on a GVF sub-dataset and tested on the remaining sub-datasets to evaluate generalizability.The setup samples eight frames from the first 32 frames of each video, center-crops and resizes them to 224 × 224, and applies augmentation.Source: A. Experimental Setup · p. 4

B. Evaluation

  • Video-level accuracy and AUC are the main evaluation metrics, with accuracy computed at a 0.5 threshold.For image-based methods, frame-level predictions are averaged into video-level predictions.Source: A. Experimental Setup · p. 4, B. Evaluation · p. 4

Q1. How does our detector generalize to unseen T2V models compared to previous detectors?

DeCoF generalizes better to unseen video-generation models than prior detectors, and remains effective under blur and JPEG compression. The paper concludes that it generalizes across open-source and commercial models.

Q1. How does our detector generalize to unseen T2V models compared to previous detectors?

Q2. How does our detector maintain robustness against noise and disturbances?

  • AUC remains above 80% under all tested degrees of Gaussian blur and JPEG compression, despite inconsistencies across training datasets.The tests use Gaussian blur at σ = 1, 2, 3 and JPEG quality levels 90, 80, 70, 60, and 50.Source: Q2. How does our detector maintain… · p. 5

VI. CONCLUSIONS AND FUTURE WORK

  • The authors conclude that DeCoF generalizes across multiple open-source and commercial closed-source generation models.Source: §VI · p. 5

APPENDIX

The appendix lists supplementary material on dataset construction, training, robustness, mapping functions, frame sampling, commercial models, spatial and temporal artifacts, limitations, and visual examples. The problem formulation defines training on one generator and evaluating generalization to unseen generators and cross-category videos.

APPENDIX

  • The appendix points readers to supplementary sections on dataset construction, model training, robustness, mapping functions, frame count and sampling, commercial models, spatial versus temporal artifacts, limitations, and visual examples.Source: APPENDIX · p. 8

1) Problem Formulation:

2) Dataset Composition:

GVF is built from real videos and prompts paired with generated videos from selected models, then assessed across prompt attributes and video quality and alignment. Its evaluation reports differences in generators’ spatial and temporal quality and attribute control.

2) Dataset Composition:

  • Prompts and real videos come from MSVD and MSR-VTT, with prompts categorized by spatial and temporal content and attributes.Source: 2) Dataset Composition · p. 8
  • The selected generators include four open-source models for the dataset and commercial models held out for secret testing, including Pika, Gen-2, Kling, Sora, and Veo.Source: 2) Dataset Composition · p. 8
  • 964 triples pair real videos and prompts with generated videos, and GVF is divided into four generator-specific subdatasets.The four subdatasets are Text2Video-zero, ModelScopeT2V, ZeroScope, and Show-1.Source: 2) Dataset Composition · p. 8

3) Dataset Evaluation:

  • The evaluation assesses static quality, temporal quality, overall video-text alignment, and fine-grained alignment to specific attributes.Source: 3) Dataset Evaluation · p. 9
  • Manual evaluation finds that people, animals, and vehicles have poorer spatial quality than plants, scenery, and natural items, with high-frequency details cited as a challenge.Source: 3) Dataset Evaluation · p. 9, Figure 9 · p. 10
  • Text2Video-zero performs well on static quality but poorly on temporal quality, while Show-1 ranks second in static quality and has better temporal quality.Source: 3) Dataset Evaluation · p. 9
  • The four models control color and camera view better than quantity, motion direction, and event order; ModelScope and ZeroScope are promising on speed and direction, while Text2Video-zero performs poorly.Source: 3) Dataset Evaluation · p. 10
  • Fine alignment is more accurate for assessing a specific attribute, while global alignment is a reasonable approximation.Source: 3) Dataset Evaluation · p. 10

B. Parameter Description

The experimental setup trains on one GVF subdataset and tests on others, using sampled and resized frames. The paper reports perturbation robustness and finds that the mapping function strongly affects performance.

B. Parameter Description

  • Training uses 771 examples, with 96 for validation and 97 for testing; CLIP:ViT is frozen, and training uses SGD.Videos are represented by eight frames, center-cropped and resized to 224×224; training also uses augmentation.Source: B. Parameter Description · p. 10
  • The model uses two transformer layers and an MLP head, with 768-dimensional features and four attention heads.Source: B. Parameter Description · p. 10

C. Robustness

  • Under Gaussian blur and JPEG compression, AUC stays above 80%; under every tested JPEG quality, it exceeds 90% for models trained on any subdataset.Source: C. Robustness · p. 11

D. Different Mapping Functions

  • The mapping function largely determines effectiveness; the CLIP image encoder is preferred over ImageNet-trained classification features for this approach.The authors state that CLIP’s image-text alignment task better supports clustering similar images, while a CLIP-task ResNet is unsuitable as a mapping function.Source: D. Different Mapping Functions · p. 11

E. The Impact of Video Frame Number

Frame-count and sampling choices affect detection performance, while the authors also test human judgment on Sora videos and whether DeCoF detects temporal artifacts without spatial cues.

E. The Impact of Video Frame Number

  • The authors selected 8 frames to accommodate current limits on long-video generation and to test their method under a challenging frame budget.They note that even videos then displayed by Sora contained about 300–600 frames.Source: E. The Impact of Video Frame Number · p. 11
  • 16 frames generally outperform 8, except for models trained on Text2Video zero, while performance worsens at 24 frames.The authors suggest longer sequences may expose more model-specific temporal artifacts and weaken generalizability.Source: E. The Impact of Video Frame Number · p. 11

F. The Impact of Video Frame Sampling Interval.

  • Across three frame-sampling strategies tested on Sora, performance showed no significant decrease, but the authors expect greater degradation for longer videos.The test videos were around 10 seconds; the authors attribute the observed robustness partly to prompts covering varied change speeds.Source: F. The Impact of Video Frame Sampling… · p. 11

2) The detection performance of human testers on Sora:

1) Detect generated videos from temporal artifacts:

  • The model detected test data without spatial artifacts and data from the earlier probe’s second set; the authors consider repeated stillness across frames a temporal artifact.Source: 1) Detect generated videos from… · p. 12

2) Spatial artifacts vs Temporal artifacts.:

The authors distinguish temporal-artifact detection from spatial-artifact detection, whose cross-model generality is difficult, and limit their method’s stated scope to generated videos. They also describe potential misuse and provide visual examples.

2) Spatial artifacts vs Temporal artifacts.:

I. Impact Statements

  • The authors warn that detection tools may be misused, potentially intensifying rivalry between video-generation and detection technologies.They also advocate ethical technology use and research on media-authenticity verification.Source: §I · pp. 12–13

J. Limitation

  • Deepfake and malicious-editing videos may cause significant performance degradation or make the method inapplicable; the method adds no face-specific detection module.The authors contrast locally tampered Deepfakes with generated videos, which are produced as a whole and span more diverse forgery scenarios.Source: J. Limitation · p. 14

K. More Visual Examples

Loading 2402.02085v8…