acceptodds
Withdrawn from ICLR 2027

Category-Oriented Universal Adversarial Perturbations for Deepfake Video Detectors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fake video detectors rely on different forensic evidence, including frame-level artifacts, pretrained visual representations, and temporal dynamics. This diversity makes it difficult for a universal adversarial perturbation to transfer across different detectors. We categorize existing detectors into three categories: Spatial-Artifact-Oriented (SAO), Semantic-Representation-Oriented (SRO), and Temporal-Dynamics-Oriented (TDO), and develop three category-specific attacks within a category-oriented framework. Specifically, we introduce frequency-aware optimization for SAO, a dual-band static perturbation for SRO, and frame-specific temporal perturbations for TDO. We further develop a unified attack that combines these complementary properties into a single structured perturbation and jointly optimizes it across the three detector categories. Experiments on four surrogate and six unseen detectors, multiple video generators, and external datasets show that the proposed ategory-specific attacks consistently improve transferability over conventional baselines. The unified attack further provides broader cross-category transfer, although with a trade-off in attack strength compared with separately optimized category-specific attacks. These results highlight the importance of detector forensic evidence in adversarial transfer and motivate robustness evaluation across different detection paradigms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fake video detection, category-oriented attacks, adversarial transferability, universal adversarial perturbations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2dTiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a category-oriented framework for universal adversarial attacks against AI-generated video detectors. It divides detectors into three categories based on spatial artifacts, semantic representations, and temporal dynamics, and designs corresponding attack strategies. Experiments show improved transferability across unseen detectors and datasets, while the unified attack achieves broader but less consistent cross-category performance.

Critical strengths
  1. A coherent category-oriented perspective. The division into SAO, SRO, and TDO according to forensic evidence offers an intuitive explanation for the limited transfer of conventional universal perturbations. The proposed attack parameterizations are aligned with this framing, making the methodological motivation easy to understand and the research question relevant to robustness evaluation.
  2. Substantial improvements on unseen detectors, supported by multiple datasets and a targeted ablation. On GenBuster-ID, the reported strict conditional Fake-to-Real flip rate increases from 14.21% to 64.61% for LGrad and from 3.90% to 85.10% for NSG-VD. The temporal-length ablation in Appendix C further shows that added frame-specific variation strongly benefits the TDO surrogate ReStraV, but not the reported spatial and semantic surrogates. These results provide useful evidence that attack behavior differs across detector categories.
Critical weaknesses
  1. Insufficient isolation of the source of the reported gains. The frame-specific attack has many more learnable degrees of freedom than a static perturbation, so the observed improvements do not by themselves establish that forensic-category alignment, rather than increased parameterization capacity or its interaction with the detector, is the main cause. The unified attack also changes perturbation structure, surrogate objectives, temporal centering, and gradient handling together, without a matched multi-surrogate ablation isolating PCGrad and centering. This weakens the causal support for the paper's central design claims.
  2. Unclear generalization protocol and limitations of practical temporal transfer. The paper should specify the optimization/evaluation video split and whether perturbations are trained separately for different datasets or transferred unchanged. Moreover, position-specific perturbations have not been tested under different frame-sampling offsets or strides, and still-frame visualizations with an ℓ∞ budget do not establish an absence of temporal flicker. The unified attack also sacrifices substantial TDO strength: on GenBuster-ID, D3 and NSG-VD achieve 21.73% and 9.04%, versus 88.10% and 85.10% with the category-specific frame-specific attack. Together, these points limit how confidently one can interpret broad universality and deployment robustness.
Questions for authors
  1. Can the authors clarify the exact video-level optimization and evaluation splits on each dataset, and whether the same universal perturbation is reused across datasets or optimized separately on each one? This distinction materially affects the generalization claim.
  2. Can the authors provide (i) a capacity-matched comparison of static and frame-specific perturbations, and (ii) a unified-attack ablation that holds the surrogate models, perturbation structure, data, budget, and schedule fixed while varying temporal centering and PCGrad? Which components explain the observed transfer gains and losses?
  3. How sensitive are the frame-specific and unified attacks to changes in the target's frame-sampling offset, stride, or sampled sequence length? Can the authors also assess temporal visual quality using continuous video playback or a temporal perceptual metric, rather than only still frames?
Reviewer_Fg14OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the Universal Adversarial Perturbation (UAP) black-box migration attack between Multiple AI-Generated Video (AIGV) Detectors. The authors argue that different detectors rely on different Forensic Evidence, which makes it difficult for traditional UAP to achieve effective cross-model transfer. To this end, the detectors are divided into three categories: Spatial-artificial-oriented (SAO), Semantic-Representation-Oriented (SRO) and Temporal-Dynamics-Oriented (TDO). Frequency-aware Perturbations, Dual-band perturbations and Frame-specific perturbations are designed respectively, and structured perturbation fusion and PCGrad are further used to implement Unified Attack. Experiments on multiple public data sets and detectors with different detection paradigms verify the attack effectiveness and migration ability of the proposed method.

Critical strengths
  1. This paper focuses on the adversarial transfer problem between heterogeneous AIGV Detectors, and analyzes the attack strategies from the perspective of the differences in forensic evidence that detectors rely on, which provides a valuable research perspective for understanding the adversarial vulnerability of different detection paradigms.

  2. Different perturbation structures are designed for spatial forgery traces, pre-trained visual representations and temporal dynamics, and cross-category unified attack is further explored. The overall framework is clear, and the method design has a good correspondence with the research motivation.

Critical weaknesses
  1. In this paper, AIGV Detectors are classified into three categories, but the basis of classification and the boundaries of the categories are not fully demonstrated. At the same time, the lack of systematic Cross-Category Attack Comparisons makes it difficult to prove that the attack performance improvement indeed comes from the matching relationship between the perturbation design and Forensic Evidence, rather than the general effectiveness of the attack strategy itself.

  2. Frequency Reweighting, Dual-band Perturbation and Frame-specific Perturbation are used for different Detector categories. And through PCGrad for combination, the overall design is more empirical, lack of unified methodology support. Moreover, different perturbation structures have different parameter degrees of freedom, and it is not clear whether the performance gain comes from Category-specific Design, or more flexible perturbation parameterization.

Questions for authors
  1. Can Cross-Category Attack Comparisons be complemented to prove that there is indeed a correspondence between different perturbation strategies and Detector Forensic Evidence, rather than that the attack strategies themselves are universally valid?

  2. With different degrees of freedom of perturbation parameters for different attack methods, can parameter-matched Comparisons be provided to rule out that the performance gain mainly comes from a larger Parameter space?

Reviewer_oE1YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper categorizes existing deepfake video detectors into three types: Spatial-Artifact-Oriented (SAO), Semantic-Representation-Oriented (SRO), and Temporal-Dynamics-Oriented (TDO). It develops a category-specific adversarial attack for each type and then introduces a unified attack that jointly optimizes a structured perturbation across the three categories using PCGrad. Experiments show that the category-specific attacks generally improve attack effectiveness over static pixel-space perturbation baselines. However, the unified attack exhibits limited transferability to unseen SRO and TDO detectors.

Critical strengths
  • Important research question. The paper presents a pioneering study of the adversarial robustness of fake video detectors.

  • Useful taxonomy of fake video detectors. The paper groups existing detectors into three categories: SAO, SRO, and TDO, based on their dominant forensic evidence, providing a clear basis for designing category-specific attacks.

  • Extensive experiments. The main attack experiments cover four datasets and ten detectors, using videos from three video generators.

Critical weaknesses
  • Limited comparisons with existing attacks. The current baselines are too simple, using the same static pixel-space perturbation across all video frames. How would attacks adapted from image-based deepfake detection, such as those in [1], perform across all detector categories? Furthermore, how would they perform against TDO detectors when perturbations are generated separately for each video frame?
    [1] Adversarial Perturbations Fool Deepfake Detectors. IJCNN 2020.

  • Insufficient empirical justification for PCGrad. The authors use PCGrad to jointly optimize the learnable parameters of the unified attack across multiple attack objectives. However, how does the unified attack perform under standard optimization without PCGrad, or with alternative optimization methods such as Nash-MTL [2] and FAMO [3]?
    [2] Multi-Task Learning as a Bargaining Game. ICML 2022.  [3] FAMO: Fast Adaptive Multitask Optimization. NeurIPS 2023.

  • Unclear transferability to unseen detector categories. New detector categories may emerge in the future, but it remains unclear how well the unified attack would transfer to them. For example, under the current setting, how would the unified attack perform if optimized using objectives from only two detector categories and then evaluated on the third?

Questions for authors

Please see the questions in the weaknesses section. Addressing following additional points and questions could help improve the paper:

  • Reporting clean detector performance. As noted in the paper (lines 430–431), the relatively low attack performance may be primarily attributable to the small number of fake videos correctly classified by the detector. Reporting detector performance alongside attack performance would help clarify this relationship.

  • Unclear hyperparameter settings for SAO attack.  How is the number of detail components (M) selected for the SAO attack and the unified attack?

  • Unclear temporal perturbation length for the experiments. Although the paper discusses the effect of temporal perturbation length and specifies the number of input frames for TDO detectors, what is the default temporal perturbation length used for the unified attack in the experiments?

  • Unclear selection criteria for white-box surrogate detectors. What criteria determine the choice of white-box surrogate detectors? For example, why are CLIP and DINOv2 both selected as SRO surrogates, while ViT-B/16 is reserved as a black-box target?

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

Related papers

Discussion (0)

Sign in to comment.