acceptodds
Withdrawn from ICLR 2027

Resource-Constrained Deployment-Time Defense Against Adversarial Observation Perturbations in Cooperative Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adversarial observation perturbations pose a critical threat to cooperative autonomous driving because corrupting the observation stream received by a single connected and automated vehicle (CAV) can induce unsafe decisions whose consequences propagate through multi-vehicle interactions. Existing defenses largely rely on robust policy training or do not explicitly consider a bounded online recovery budget, limiting their applicability when a pretrained driving policy must remain unchanged and only a small number of observations can be repaired before each decision. We study this deployment-time problem under a localized compromise in which one victim CAV receives persistently corrupted observations while its identity is unknown to the defender. We propose Multi-view Evidence-guided Localization and Locking with Resource Allocation and World-Model Recovery (MELAR) for this setting. MELAR uses cross-view consistency anchored by protected ego states to localize a suspected victim CAV, abstaining when the evidence is insufficient. It then ranks surrounding-vehicle observation blocks according to corruption evidence and traffic-interaction risk, and reconstructs only the highest-priority blocks within the available per-step budget. A dual-scale gated recurrent unit (GRU) world model integrates temporal cross-view evidence to reconstruct the selected blocks before the restored observations are passed to the frozen driving policy. Under the evaluated white-box observation attack, results averaged over three evaluation seeds show that, with a per-step recovery budget of three observation blocks, MELAR reduces the collision rate from 64.0% for the undefended policy to 30.7%, a reduction of 33.3 percentage points. These results show improved closed-loop driving safety in the evaluated setting while keeping the deployed policy fixed.

Primary area:
applications to robotics, autonomy, planning
Keywords:
cooperative autonomous driving, adversarial robustness, observation perturbations, deployment-time defense, resource-constrained recovery, world models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CjrsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to repair corrupted observations without changing the deployed driving policy. It identifies the attacked vehicle through comparisons across vehicles, then selects a small number of observation blocks and reconstructs them from other vehicles’ observation histories.

Critical strengths

The paper addresses a practical deployment problem without retraining the driving policy, and its closed loop experiments show a substantial reduction in collision rate under the tested attack.

Critical weaknesses
  1. The AI generated figures appear insufficiently reviewed, with visible layout errors and a ranking formula in Fig. 2 that conflicts with Eq. 7.

  2. The defense may not activate against background vehicle attacks permitted by its own threat model, and its effectiveness against adaptive attackers remains untested.

Questions for authors
  1. I do not object to using AI tools to create figures, but the authors appear to have made no effort to review or revise the AI-generated images. The third panel of Fig. 2 has careless layout, with text extending beyond its boxes, overlapping components, and even garbled symbols. There are also unexplained gray dashed lines in the second panel of Fig. 1. More importantly, the ranking formula in the second panel of Fig. 2 is inconsistent with Eq. 7. In my view, figures of this quality do not meet the presentation standard expected of an ICLR submission.

  2. Section 3 explicitly allows the attacker to modify only background-vehicle observation blocks, while Section 4.1 acknowledges that such modifications may leave the protected ego-anchor residuals unchanged. If the localizer abstains, no recovery query is issued. An attacker could therefore target background vehicles relevant to the driving decision while bypassing the entry point to the defense. The paper does not evaluate this attack, even though it falls within its stated threat model.

  3. The experiments do not adequately test a white-box attacker targeting the defense itself. An attacker who knows the defense could avoid its anchor checks or keep important blocks out of the top KK recovery targets. The experiments therefore do not show whether MELAR can withstand an adaptive attack.

  4. KK limits only the number of blocks selected for recovery at each step. The system makes a decision every 0.05 seconds and uses cross-view evidence delayed by 0.25 seconds, but the paper reports no runtime, missed deadlines, or time from attack onset to recovery.

  5. The paper mainly reports final collision rates. It does not clearly report how often the detector locks onto the wrong vehicle without an attack, its localization accuracy and delay after an attack, what fraction of selected blocks were actually corrupted, or how often a recovery query results in a writeback.

  6. I also have concerns about the writing. The paper does not clearly distinguish how it uses cooperative and collaborative. MAPPO and MARL should be expanded and briefly explained when they first appear in the problem formulation. Although symbols such as b\mathcal{b} and MM are used, I still found it difficult to understand what they represent in concrete terms.

Reviewer_98s3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper formulates a problem of deployment-time observation restoration: under conditions where the driving policy is unchanged, the victim vehicle is unknown, and the per-step block-query budget is limited, how can cross-vehicle information be used to restore adversarially perturbed observations? To address this, the authors propose MELAR, which combines source localization, target selection, and dual-scale state reconstruction.

Critical strengths
  1. The paper focuses on the impact of single-vehicle observation attacks on cooperative driving and clearly defines the research conditions: a frozen policy, an unknown victim, and a limited repair budget.
  2. The method’s structure is clear. Source localization, target selection, and temporal reconstruction basically correspond to the difficulties identified earlier.
Critical weaknesses
  1. The single-vehicle attack setup is acceptable, but the authors should explain why only one vehicle’s observations of surrounding vehicles are attacked, while other vehicles’ observations, as well as the ego state and identity information of the victim vehicle, remain protected. What real-world scenario does this setting correspond to?
  2. The authors explain the burden of retraining policies, but which practical scenarios require keeping the original policy unchanged and why the number of reconstruction queries per step is limited still need clarification, especially the relationship between the query budget and actual computational cost.
  3. The novelty of the method is not clear enough. The paper combines cross-vehicle consistency detection, risk ranking, and a dual-scale GRU, but the authors still need to explain which technical contributions, beyond the problem setup, distinguish it from existing methods.
  4. It’s unclear whether the localization is accurate. Direct results are missing. Removing localization shows a performance drop, suggesting it helps, but it’s still unknown whether it can identify the victim vehicle promptly and how often it identifies the wrong vehicle.
  5. The extra benefit of dual-scale reconstruction is small. The collision rates for direct replacement and single-scale GRU are both 32.0%, while the full method is 30.7% in Table 2. Currently, this is insufficient to show whether the dual-scale design is consistently better, particularly whether it handles information delay better, which would require comparisons with these alternatives across different observation delays.
Questions for authors

What practical attack scenario supports the assumed separation between vulnerable surrounding-vehicle observations and protected ego states, identity information, and other vehicles’ observations? Which deployment requirements motivate keeping the policy unchanged?

Reviewer_Yk7sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MELAR, a deployment-time defense that repairs corrupted observations while keeping the driving policy frozen. It combines protected-anchor victim localization, risk-aware repair allocation, and dual-scale GRU reconstruction.

Critical strengths

The paper addresses a practical setting with a frozen policy, unknown victim identity, and a limited repair budget. It demonstrates substantial improvements in closed-loop collision rates while clearly acknowledging its assumptions, evaluation limitations, and modest component-level gains.

Critical weaknesses

W1. The evaluation does not establish robustness against attacks adapted to the complete defense.## Section 5.1 describes PGD through the frozen actor, without optimizing against MELAR’s localization or repair mechanism. Sections 3 and 4.1 also acknowledge that background-vehicle-only attacks may evade localization, but this permitted attack is not evaluated.

W2. Episode-long source locking makes early attribution errors difficult to recover from.## Section 4.1 specify that the selected source remains locked until episode termination, even when another source later receives a higher diagnostic score. The experiments do not quantify the consequences of incorrect locks.

W3. The additional benefit of the proposed components is not convincingly established.## In Table 2, full MELAR reduces the collision rate by only 1.3 percentage points, approximately 4.1% relative to Direct Proxy. This raises concerns about whether the modest gain reliably justifies the added complexity.

W4. MELAR’s impact on clean driving performance is not established by the reported experiments.## Table 1 includes Clean MAPPO but provides no corresponding evaluation with MELAR enabled in the absence of attacks. Although the detector is calibrated on clean trajectories, it remains unclear whether false alarms and unnecessary repairs could degrade normal driving performance.

Questions for authors

1.How does MELAR perform against defense-aware attacks, particularly background-vehicle-only perturbations that may evade localization?

2.Does MELAR’s modest advantage over Direct Proxy persist across additional random seeds and evaluation conditions?

3.With MELAR enabled but no attacks present, what are the collision rate and false-activation rate compared with Clean MAPPO?

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

Related papers

Discussion (0)

Sign in to comment.