acceptodds
Under review as a conference paper at ICLR 2027

Beyond Generator-Specific Shortcuts: Open-World AI-Generated Video Detection via Adversarial Training

Abstract

AI-generated video (AIGV) detection is increasingly important for mitigating misinformation, impersonation, and other malicious uses of synthetic media, yet existing detectors often generalize poorly to unseen generators in open-world settings. To better understand this challenge, we analyze the failure modes of AIGV detectors under cross-generator out-of-distribution (OOD) shifts and observe that detector representations strongly encode generator-specific cues, which are associated with substantial ID–OOD performance drops. Further analysis shows that naively perturbing generator-sensitive tokens can even reduce OOD performance, indicating that generator-related cues and detection-relevant evidence are only partially separable. Motivated by these findings, we propose Forensic Obfuscation of geneRator-specific cues via Gradient-guided Embedding perturbation (FORGE), a token-aware adversarial training framework for open-world AIGV detection. FORGE identifies tokens that are strongly associated with generator-specific shortcuts yet less critical to real/fake discrimination, and selectively perturbs them during training to suppress generator-dependent patterns while preserving useful forensic evidence. Through synchronized adversarial training between an attacker and a detector, FORGE encourages the detector to learn more transferable generator-invariant forensic representations. Experiments on 29 held-out benchmark entries show that FORGE achieves an average F1 of 92.62, outperforming eight strong baselines by 5.31 points, while also remaining robust under four common post-processing attacks. We will release our code and datasets upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Video Detection

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_wmf6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes Forensic Obfuscation of geneRator-specific cues via Gradient-guided Embedding perturbation (FORGE), a token-aware adversarial training framework for open-world AIGV detection. FORGE identifies tokens that are strongly associated with generator- specific shortcuts yet less critical to real/fake discrimination, and selectively perturbs them during training to suppress generator-dependent patterns while preserving useful forensic evidence. Extensive experiments validate the effectiveness of the proposed method.

Critical strengths

1.The basic idea is good. The paper targets a practically important problem—generator-specific shortcut reliance in open-world AIGV detection—and addresses it with a coherent token-aware adversarial training framework. The selective perturbation criterion is well motivated, and the intervention analysis provides direct empirical support for the core premise.

2.The performance compared with SOTA is good. FORGE achieves 92.62 average F1 on 29 held-out entries, outperforming eight representative baselines by 5.31 points, with consistent gains across backbones, informative ablations, robustness under post-processing and white-box attacks, and a family-disjoint subset analysis that strengthens the empirical claims.

Critical weaknesses
  1. The "generator-sensitive but detection-safe" dichotomy lacks theoretical guarantee and may not hold universally. The core assumption of FORGE is that there exists a class of tokens strongly encoding generator identity yet contributing little to real/fake discrimination, so perturbing them removes shortcuts without harming detection. This is supported only empirically, with no theoretical justification. If a new generator's artifacts are entangled with genuine forensic cues—e.g., its compression fingerprints are also key to real/fake discrimination—the assumption breaks down, and selective perturbation could damage detection performance. A discussion of boundary conditions, or a stress test where generator-specific and detection-relevant cues are deliberately correlated, would strengthen the claim.

  2. The token importance measure relies on a classic gradient-norm heuristic. While computationally efficient, it is not compared against more recent, faithful ViT attribution methods. As a result, it remains unclear whether the observed gains stem from the proposed selective perturbation strategy or simply from the choice of a coarse importance proxy, and whether a more faithful measure would further improve OOD generalization.

  3. The "open-world" claim is not fully supported by the experimental setup. Although the paper evaluates on 29 held-out entries, all videos are uniformly sampled to 8 frames and resized to 224×224, and are drawn from only three public benchmarks. This setup primarily tests cross-generator generalization rather than true open-world conditions, which would involve longer videos, higher resolutions, multimodal inputs, and repeated editing or compression. The paper acknowledges in its limitations that longer videos and broader real-world domains remain uncovered, but the title and abstract still frame the contribution as "open-world" generalization. This may be a cross-generator OOD study rather than an open-world one. A more precise framing or additional evidence beyond the current benchmark would make the claim more defensible.

Questions for authors
  1. When generator-specific artifacts are strongly correlated with genuine forensic cues, can the proposed method still select the right tokens? Is there a stress test for this entangled case?
  2. The perturbation is guided solely by the generator classifier's KL gradient. If the generator classifier is biased, could the perturbation direction be misled? Is there a control using the real/fake gradient to guide perturbation?
  3. How do you delineate "open-world" from "cross-generator OOD"? Do the conclusions hold on longer videos or realistically compressed content?
Reviewer_EUqxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FORGE, a training framework for AI-generated video detectors addressing generalization to generators unseen during training. It employes adversarial training to perturb tokens that are mostly associated with specific generators, while preserving the ones that are useful for discriminating between real and generated videos, applying a sort of regularization to the trained detector.

Critical strengths
  • The addressed problem is relevant and timely. It is linked to practical applications yet remains an open challenge.

  • The provided insights on generalization degradation and token selection are interesting and clearly motivate the proposed approach.

Critical weaknesses
  • The overall novelty is not clear, and might be limited. In particular, the motivations and main principles are very similar to those proposed in [a]. Also [b] seems to be based on similar findings. The authors should discuss how their work relates to them. Moreover, the novel contributions that appear are not properly described or justified. For instance, in the token selection rule, it is not clear how gradients are backpropagated to individual tokens, since this strictly depends on the model architecture. Furthermore, the ablation study seems to miss important settings, e.g., score-based selection with random perturbation, sensitivity to the hyperparameters, and to the chosen/number of training generators.

  • The considered experimental setting might overestimate the "open-world" generalization capabilities, as 11 of the 29 held-out generators are from the same families of training ones. The considered baselines are arguable: [b] has shown that fully fine-tuning CLIP affects generalization; there are no baselines with a linear probe, neither employing standard adversarial training on real/fake loss or simple regularization techniques like common data augmentations; UNITE is not deployed with its optimal parameters (it has been shown to be less effective with a small number of frames); none of the competing methods is designed for robustness. Moreover, the authors do not report the false positive rate or threshold-agnostic metrics such as ROC-AUC, and do not repeat the experiments with different random seeds (reporting only the average and standard deviation). Finally, the validation set, the post-processing transformation settings, and the complete hyperparameters of the adversarial attacks are not described.

[a] Yan, Z., Zhang, Y., Fan, Y., & Wu, B. (2023). UCF: Uncovering Common Features for Generalizable Deepfake Detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 22355-22366.

[b] Yan, Z., Wang, J., Jin, P., Zhang, K., Liu, C., Chen, S., Yao, T., Ding, S., Wu, B., & Yuan, L. (2024). Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection. International Conference on Machine Learning.

Questions for authors
  • Could you please better describe in the paper what a token is in your setting, and the complete architecture of the models (in particular, the layers on top of the CLIP backbone)?
  • Is it possible that using a single source of real videos introduces a bias, as models might learn to distinguish the datasets?
  • Are real videos perturbed? If not, could the perturbations serve as a cue for fake videos?
  • Could you please report training time, memory costs, and runtime of FORGE with respect to the CLIP baseline?
Reviewer_enhkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes FORGE, a token-aware adversarial training framework for improving the generalization of AI-generated video detectors to unseen generators. The authors observe that existing detectors tend to encode generator-specific information, which may contribute to poor out-of-distribution generalization. To address this issue, FORGE employs an attacker–detector training framework. The attacker identifies tokens that are highly relevant to generator identification but relatively less important for real/fake discrimination, and applies gradient-guided perturbations to their representations. The detector is then optimized on clean and perturbed representations using classification and consistency objectives to reduce reliance on generator-specific shortcuts. Experiments across 29 held-out evaluation entries demonstrate improved detection performance over eight baselines. Additional ablation studies and robustness evaluations support the effectiveness of the proposed framework.

Critical strengths

Relevant and well-motivated problem. Improving generalization to unseen video generators is an important challenge in AI-generated video detection. The paper provides a reasonable motivation by investigating the generator-specific information encoded in detector representations.

Interesting method design. The joint consideration of generator relevance and detection relevance for token selection is intuitive and meaningful. Unlike simply suppressing generator-related information, the method attempts to preserve features important for real/fake discrimination.

Strong empirical performance. FORGE achieves substantial improvements across the evaluated OOD datasets, including a 5.31-point average F1 improvement over the strongest baseline. The additional family-disjoint evaluation further strengthens the experimental evidence.

Critical weaknesses

Insufficient analysis of adversarial training dynamics and sensitivity to training duration.

FORGE relies on alternating optimization between an attacker and a detector, where the attacker continuously adapts its token selection to the evolving detector representations. However, the paper provides limited evidence regarding the stability of this process.

The experiments use a fixed 12-epoch training schedule, but it remains unclear how sensitive the final performance is to the number of training epochs or whether the attacker and detector objectives exhibit stable convergence. Since the effectiveness of the proposed framework depends on their interaction, analyzing training dynamics, such as validation performance and token-selection stability across epochs, would help establish the reliability of the method.

The current evaluation may not sufficiently challenge open-world AIGV detection.

Although the paper evaluates 29 held-out entries, several relatively simple or established baselines already achieve strong performance on many of them. For example, the supervised CLIP baseline and DeMamba obtain average OOD F1 scores of 85.62% and 86.93%, respectively, with nearly saturated performance on several generators.

This raises concerns about whether the current benchmark adequately reflects the difficulty of detecting high-quality, realistic AI-generated videos in challenging open-world settings. Evaluation on additional challenging benchmarks, such as MintVid and Artifact-Bench, would provide stronger evidence for the practical generalization capability of FORGE.

Sensitivity to the number of training generators.

FORGE relies on an auxiliary generator-ID classifier to identify generator-specific tokens, making the diversity of training generators an important factor in the proposed framework. However, all experiments use a fixed set of 10 training generators. It remains unclear how the effectiveness of the proposed token selection and the resulting OOD generalization depend on the number of available training generators. An ablation varying the number of training generators, while controlling the training data volume where possible, would help establish the robustness and applicability of the method under different training conditions.

Questions for authors

Training dynamics: Could the authors clarify how the 12-epoch training schedule was selected? Do the existing training logs show stable validation performance and attacker/detector losses across epochs?

Sensitivity to training duration and the number of training generators: How sensitive is FORGE to the number of training epochs, and does its performance remain stable without careful early stopping? Furthermore, given its reliance on generator-ID supervision, how does the number and diversity of training generators affect OOD generalization? Does FORGE maintain consistent improvements when fewer generator identities are available?

Benchmark difficulty: Could the authors comment on the generalization of FORGE to more challenging datasets, such as MintVid or Artifact-Bench, particularly given the strong performance of simple baselines on the current evaluation?

Reviewer_dnXoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the open‑world AI‑generated video (AIGV) detection problem, where existing detectors heavily learn generator‑specific shortcut cues and suffer severe performance drop on unseen video generators. It uses gradients from generator-classification and real/fake detection objectives to identify tokens that are informative about generator identity but less important for detection. The detector is trained on both clean and perturbed representations, optimized with real‑fake detection loss and forensic consistency loss to learn generator‑invariant forensic features. Then, experiments conducted across benchmark entries illustrate its effectiveness.

Critical strengths
  1. The analysis about why many prior detectors fail under cross‑generator OOD setting is valuable for open‑world synthetic video forensics research.

  2. It evaluates multiple held-out entries and detectors, and includes a family-overlap screening. The evaluation is relatively comprehensive.

Critical weaknesses
  1. It does not analyze concrete failure samples. It is unclear whether failure cases stem from remaining generator shortcut artifacts.

  2. The main results do not provide variability across independent training runs or confidence intervals. Some entries also have imbalanced class counts after integrity checks, and sample sizes may vary across generators.

  3. The main OOD benchmark is not fully generator-family-disjoint.

  4. All real videos are sampled from Vript, whereas generated videos come from three public repositories. The real/fake data-source setup leaves a substantial confounding concern.

Questions for authors
  1. All evaluation uses fixed 8‑frame sampling. Could you add small‑scale results to show performance when feeding full‑length videos or different number of sampled frames?

  2. How are classification thresholds selected? Are they determined using only the ID validation set and then held fixed across all OOD entries and baselines?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.