acceptodds
Under review as a conference paper at ICLR 2027

PACE: Persistent Adversarial Coefficient Evolution for Recurrent Visual Agents

Abstract

Recurrent visual agents maintain internal states that integrate past observations, reducing the immediate impact of isolated perturbations while introducing a new attack surface: persistent adversarial influence over their evolving internal dynamics. Existing attacks typically optimize each frame independently, ignoring the temporal dependency introduced by recurrent policies and resulting in ineffective or unstable perturbations. We propose Persistent Adversarial Coefficient Evolution (PACE), a framework that formulates adversarial manipulation as the evolution of a compact latent attack state over time. PACE first constructs a low-dimensional adversarial basis by querying the frozen agent, generating performance-degrading perturbations, and extracting dominant perturbation directions through factorization. During interaction, the previous attack coefficient initializes the current optimization, while temporal smoothing regulates abrupt changes in the adversarial state. The resulting perturbations are projected onto the valid constraint and image domain before being applied to the victim agent. Through warm-started coefficient updates and temporal smoothing, PACE retains attack information across frames while adapting perturbations to the current observation. Experiments on DreamerV3 across DMC Walker Walk, Atari Pong, and Crafter show that PACE achieves the largest reported mean reward reduction among the compared methods, tying with UAP-RL on Crafter, while exhibiting substantially lower temporal variation than MAD, PA-AD, Illusory, and DAPGD. These results highlight the importance of modeling adversarial attacks as evolving processes when evaluating the robustness of recurrent visual decision-making systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
visual agents, recurrent policies, adversarial attacks, temporal robustness

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

The paper proposes PACE, a persistent adversarial attack framework for recurrent visual agents. It first constructs a low-dimensional adversarial basis from critic-guided perturbations using SVD. During online interaction, the attack coefficients are initialized from their previous values, refined using current-observation gradients, and smoothed through an exponential moving average before being projected onto the valid perturbation set. The proposed framework aims to preserve useful adversarial information across consecutive observations while maintaining attack effectiveness and temporal consistency. Experiments on DreamerV3 across DMC Walker Walk, Atari Pong, and Crafter compare PACE with five existing attack methods.

Critical strengths
  1. This is a meaningful problem formulation for recurrent visual agents. The paper investigates an important aspect of adversarial robustness by explicitly considering the persistence of attack information across sequential observations. This perspective is relevant to recurrent world-model agents, where current decisions depend on accumulated historical context.

  2. PACE achieves competitive reward reductions on DMC Walker Walk, Atari Pong, and Crafter (Tables 1, 2 and Table 5). In particular, it obtains an 85.71% reward drop on Pong while maintaining substantially lower temporal variation than several competing attacks.

Critical weaknesses
  1. The proposed framework primarily combines current techniques, including SVD-based adversarial subspace construction, warm-started gradient optimization, and exponential moving average smoothing. Although applying these techniques to recurrent visual agents is maybe reasonable, PACE does not introduce a substantially new optimization objective or mechanism that explicitly models long-term recurrent state dynamics. This is not very interesting to some extent. Because the temporal persistence is mainly implemented by reusing previous attack coefficients and applying standard EMA smoothing. Consequently, the technical contribution seems largely incremental compared with conventional low-dimensional adversarial optimization.

  2. The paper emphasizes maintaining attack optimization states across consecutive observations as its central contribution (Lines177–194). However, the ablation studies Table3 and table4 do not independently evaluate this mechanism. Specifically, V3 removes EMA while retaining warm-started optimization, whereas V2 simultaneously removes the subspace constraint and temporal smoothing. There is no experiment that directly compares PACE with independently initialized frame-wise optimization using the same critic-guided basis and optimization budget. Therefore, the reported improvements cannot be clearly attributed to persistent coefficient evolution, leaving the central methodological claim insufficiently validated.

Questions for authors
  1. What are the fundamental technical differences between PACE and conventional warm-started adversarial optimization in a fixed low-dimensional subspace? Could the authors clarify which component constitutes the primary algorithmic innovation beyond combining existing techniques?

  2. Could the authors provide a controlled ablation that removes coefficient warm-starting while preserving the critic-guided basis and online optimization budget? This would help determine whether maintaining the previous attack state provides an independent performance advantage.

Reviewer_pptdOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Persistent Adversarial Coefficient Evolution (PACE), a framework for persistent adversarial attacks against recurrent visual agents that combines a low-dimensional adversarial basis with warm-started coefficient updates and temporal smoothing to maintain attack consistency across frames. Experiments on DreamerV3 across multiple environments demonstrate that PACE effectively reduces cumulative rewards while substantially reducing temporal perturbation variation.

Critical strengths

The paper presents an interesting perspective by emphasizing the temporal continuity of perturbations in adversarial attacks.

Critical weaknesses
  1. Although the method is clearly described, the overall structure and narrative are not yet up to the standard of a mature academic paper.

  2. In terms of experiments, relevant analyses are lacking, such as comprehensive baseline comparisons and an assessment of the generalizability of the proposed method.

  3. The motivation needs to be further strengthened. Besides improving attack effectiveness, does considering the temporal continuity of perturbations have practical value? For example, does it make attacks easier to implement, facilitate their implementation in the physical world, or help improve defense mechanisms?

Please refer to the specific questions for details.

Questions for authors
  1. In PACE, the perturbations generated by the attacker exhibit correlations across frames. How is this correlation manifested? Can it be quantified using a specific metric to verify that PACE produces more highly correlated perturbations than methods such as PA-AD?

  2. How does PACE compare with single-step gradient-based attack methods such as FGSM and PGD?

  3. The baseline comparisons lack methods published in top-tier academic conferences and journals in 2025 and 2026.

  4. The paper's presentation is not sufficiently clear, making its contributions and novelty difficult to identify. For example, each paragraph in the Introduction and Related Work sections is only a few sentences long, which hurts readability.

  5. The experiments are conducted exclusively on DreamerV3. Is the proposed method specifically designed for DreamerV3, and how applicable is it to other world models? Can the proposed method be transferred to the RL algorithm architectures discussed in the PA-AD paper, such as PPO and DQN, including tasks that do not involve image-based control?

  6. Can this work be extended to a collaborative attacker–defender co-evolution framework to improve model robustness?

Reviewer_a6LLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

PACE attacks frozen recurrent visual agents by constructing a low-dimensional basis from offline critic-guided perturbations and maintaining online coefficients through warm-started refinement, EMA smoothing and projection. The main experiments attack DreamerV3 on Walker Walk, Pong and Crafter with an l-infinity budget of 24/255. PACE has the strongest reported mean return reduction on Walker and Pong and ties UAP-RL on Crafter, while offering a favorable temporal-variation trade-off. Supplementary experiments examine basis reuse, AI-generated technical reference | Verify before use 4 implementation-specific latency and DMC hyperparameters.

Critical strengths

The method is straightforward and the distinction between recurrent-context conditioning and persistent attacker state is useful. The threat model clearly limits deployment to current observations and fixed incoming recurrent context. The evaluation considers actual episodic performance as well as action divergence, smoothness and SSIM, and acknowledges that these metrics need not agree. The mathematical appendix carefully avoids claiming that SVD reconstruction, expected critic-bin index or EMA imply return-reduction guarantees. The supplementary cross-checkpoint and cross-architecture basis-reuse experiments are useful, while correctly retaining their target-gradient-access qualification.

Critical weaknesses

-- Core idea and causal logic: the benefit of persistence is not isolated. The central motivation is that carrying attack state across frames is particularly useful against recurrent agents. However, V3 removes EMA while retaining warm starts, and V2 changes both the search space and smoothing. No ablation resets coefficients every frame while preserving the learned basis, objective, update budget, and smoothing protocol. Appendix A.8 acknowledges this omission. Consequently, the results do not establish whether the improvement comes from cross-frame state, a better critic surrogate, or a useful fixed basis. Moreover, Action-KL conditions both predictions on the same attacked recurrent context, so it does not measure accumulated corruption of memory.

-- Algorithmic attribution: coefficient normalization changes the effective update scale. The paper uses RMS-normalized gradients with a numerical step size of 6 in both pixel and coefficient coordinates. These are not comparable pixel-space steps.

-- Experimental soundness and presentation: the baseline implementations are insufficiently specified. The shared perturbation budget and white-box access do not establish a fair optimization comparison.

Questions for authors
  • What is the effect of removing warm starts alone, and of adding persistence and EMA to the full-pixel control?
  • Are the principal rankings stable across evaluation seeds and victim checkpoints?
  • How were every baseline's optimization and offline preparation configured and tuned?
  • Why is UAP-RL's TempAbs zero under the stated onset-inclusive, post-projection definition?
  • Does a decoded-value surrogate or a smaller perturbation budget preserve the conclusions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.