acceptodds
Under review as a conference paper at ICLR 2027

SIPA: Stealthy Intent Planning Attacks on V2X End-to-End Autonomous Driving

Abstract

In vehicle-to-everything (V2X) end-to-end driving, the ego vehicle fuses bird’s- eye-view (BEV) features shared by collaborators and uses the fused represen- tation for both cooperative perception and planning. Existing V2X attacks and defenses, however, largely focus on the perception side. In particular, perception- level defenses can flag collaborators whose transmitted features visibly corrupt detection, motivating a subtler threat: can a malicious collaborator disrupt down- stream planning while keeping cooperative perception close to its clean behavior? We study this perception-stealthy planning attack and propose SIPA (Stealthy Intent Planning Attack). A key challenge is that the attacker cannot access or query the victim planner and does not observe its private route intent. Our key observation is that this intent is not entirely hidden. Planning-aware V2X proto- cols broadcast a request map generated from provisional waypoints, whose spatial structure retains a coarse proxy of the ego vehicle’s route intent. SIPA exploits this protocol-visible side channel to condition an independently trained surrogate plan- ner, infer a counter-intent steering direction, and localize a sparse request-guided perturbation region. Rather than maximizing generic waypoint displacement, SIPA optimizes a control-aware planning objective over trajectory-derived quan- tities used by the downstream controller, while suppressing perception drift. In closed-loop CARLA evaluation, SIPA reduces the Driving Score by 19.91 points while keeping cooperative perception close to the same-frame clean counterfac- tual (∆[email protected] = −0.002). As a result, a perception-level defense (CP-Guard) flags the SIPA collaborator on 18.4% of frames, at or below its 19.2% false-alarm rate on an honest one. These results reveal that perception-level consistency alone does not guarantee downstream planning safety.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Attack, V2X-E2E Autonomous Driving

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 4 reviews, above 27% of the others.
Reviewer_by3aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the question of whether a malicious collaborator can disrupt downstream path planning while maintaining a perception close to its normal behavior, this paper investigates a perception-stealthy path planning attack and proposes Stealthy Intent Planning Attack (SIPA). The key challenge lies in the attacker's inability to access or query the victim's planner or observe its private route intent. A core observation of this paper is that the victim's planning intent is not completely hidden. A path-planning-aware V2X protocol broadcasts a request map generated from temporary waypoints, the spatial structure of which retains a coarse representation of the vehicle's route intent. SIPA utilizes this protocol-visible side channel to conditionally configure an independently trained agent planner, derive the counter-intent direction, and locate the perturbation region of sparse request guidance. SIPA optimizes a control-aware planning objective for the trajectory derivatives used by the downstream controller while suppressing perception drift. In closed-loop Carla evaluation, SIPA reduces the driving score by 19.91 points while making the collaborative perception result close to the clean counterfactual result in the same frame.

Critical strengths
  1. The planning-aware request map introduced to improve communication efficiency can leak coarse proxy information about the vehicle's route intention through the protocol-visible side channel. Thus, this paper reveals a new V2X end-to-end attack surface, namely that a single malicious collaborator can severely disrupt the downstream planning process, while the collaborative awareness behavior remains close to normal.
  2. This paper proposes SIPA, which recovers the route intent without accessing the victim planner or its private route objective, and uses this intent to constrain an independently trained agent planner, locate a sparse request-guided perturbation region, and optimize a control-aware objective function on the trajectory-related quantities actually read by the deployed controller.
  3. In closed-loop Carla, SIPA significantly reduces downstream planning and closed-loop driving performance while retaining cooperative capabilities.
Critical weaknesses
  1. SIPA obtains planning gradients from an independently trained and frozen alternative planner. The authors do not clearly explain how this alternative planner was trained.
  2. In Section 3.1, the attacker can only control a single cooperative vehicle and has no access to the victim planner. However, it is also assumed that the attacker has access to the real Ego's perception input, as well as the context tensor required to run the proxy planner and perception filter. Is this a contradiction? Furthermore, how can the attacker, who controls the cooperative vehicle, obtain the real Ego's perception input and context tensor?
  3. In the results presented in Table 1, the SIPA method appears to be more balanced in terms of attack, but it does not achieve the best results in any metric, especially in the planning-related metrics, where it differs greatly from methods such as PGD. The authors sacrificed attack performance for stealth, which I believe is not an ideal outcome.
  4. This paper lists a large number of hyperparameters used in the appendix. How were these parameters determined? Were they determined through extensive experiments, or are they based on theoretical principles? It is recommended that the authors supplement the experiment with sensitivity analysis of the hyperparameters to better demonstrate the impact of each parameter on attack performance.
  5. The experiments in this paper are relatively weak. The ablation experiments did not ablate important modules, such as the Request Map. It is also recommended to supplement robustness and generalization experiments, such as the number of attackers, scenario types, different fusion models or planners, etc.
Questions for authors

Please refer to the Critical Weaknesses discussed above.

2 comments

We thank the reviewer for the comments. We respond to each weakness below.

W1 — The training of the surrogate planner is not clearly explained

We will specify the surrogate planner's training procedure in the main text (currently Appendix A.7). The surrogate is trained from random initialization on Towns 1–4 and 6, with Town05 held out, and it never loads the victim planner's checkpoint at any stage. As the reviewer notes, this explanation was in a less visible location, so we will move it to the main text and describe it more clearly.

W2 — Controlling only a single collaborative vehicle yet assuming access to the real ego's inputs seems contradictory

We apologize for the confusion. This is not a contradiction, because access is partitioned by module, not by vehicle—the perception stack is a shared implementation and can be reproduced by the attacker, whereas the victim planner is not accessible.

More fundamentally, we accepted the reviewer's point and made the threat model itself more realistic for re-evaluation. Previously we assumed the attacker could access the real ego's perception inputs and past historical values when evaluating SIPA and all baselines; agreeing with the reviewer, we re-evaluated SIPA and all baselines in a setting that does not use that information. Concretely, we use none of the real ego inputs or context tensors (including past temporal state), and we do not project the attacker's features into the victim frame to inject an estimated input. Whereas SIPA attains ΔDS −19.91 under the original setting, it is maintained with little change (ΔDS −14.90) under this setting. In other words, the very assumption of "how a collaborating-vehicle attacker obtains the real ego's perception inputs and context tensors" is not needed for the attack.

W3 — SIPA is not best on any metric and seems to sacrifice attack performance for stealth

We respectfully offer a slightly different interpretation. PGD* and SOMBRA* show larger raw ΔDS because they heavily corrupt perception, and as a result they are rejected by a perception-level defense on 71.3% and 61.4% of frames, respectively, so their effect is strongly limited on a defended system. In this light, SIPA's stealth is better viewed not as a mere sacrifice of attack performance but as the property that allows the attack to persist even in the presence of such a defense (rejection rate 18.4%, at or below the clean false-alarm rate of 19.2%). That is, the design targets effective threat under realistic, defended conditions rather than maximal harm under an idealized, undefended one.

W4 — How were the many hyperparameters determined? Sensitivity analysis is recommended

All hyperparameters of our method were set based on preliminary experiments and standard settings (the default configurations of the attack/defense frameworks we build on), and the full list and values are provided in the Appendix (ε\varepsilon, KK, α\alpha, μ\mu, λ\lambda, vmax⁡v_{\max}, τR\tau_R, τp\tau_p, etc.). For the requested sensitivity analysis, we added a sweep over the perturbation budget ε\varepsilon, the single hyperparameter that most directly governs the balance between attack strength and stealthiness. As shown below, as ε\varepsilon increases the attack strength (ΔDS) grows monotonically, but perception preservation degrades sharply beyond ε≈6\varepsilon \approx 6. Thus ε≈6\varepsilon \approx 6 is the turning point that secures attack strength while preserving perception, and the ε=6\varepsilon = 6 used in the paper is a reasonable point in this damage–stealth balance.

ε\varepsilonΔ\DeltaDSΔ\DeltaTP%Δ\DeltaFP/fΔ\Delta[email protected]Precision%
0.5−6.5-6.5−0.66-0.66−0.048-0.048+0.0016+0.001698.66
2−8.1-8.1−1.08-1.08−0.013-0.013+0.0013+0.001398.56
4−12.9-12.9−3.23-3.23−0.048-0.048+0.0009+0.000998.14
6−19.91-19.91−4.13-4.13+0.097+0.097+0.0001+0.000194.98
12−31.1-31.1−5.85-5.85+0.290+0.290−0.0436-0.043687.71
24−32.5-32.5−4.67-4.67+0.629+0.629−0.0504-0.050485.96

W5 — Experiments are relatively weak; no Request Map ablation; add robustness/generalization

We strengthen the evaluation in several ways. (1) First, an ablation of the request map is in fact already included: the random-ROI (vs. request-guided ROI) comparison and the w/o-intent-conditioning setting isolate, respectively, the contributions of the request map's spatial request and of the recovered intent. Since this was not made sufficiently clear, we will revise the labeling and description so that these two are evident as a "Request Map ablation." (2) For generalization to another fusion model, we added results on V2X-ViT. As shown below, on V2X-ViT as well SIPA substantially degrades Driving Score (ΔDS −11.19) while largely preserving perception (ΔmAP −0.0156), showing that the attack is not confined to CoDriving's specific architecture and transfers to other cooperative-perception models.

MetricCoDriving (SIPA)V2X-ViT (SIPA)
Clean →\to attacked DS83.23 →\to 68.0246.86 →\to 35.67
Δ\DeltaDS−15.21-15.21−11.19-11.19
Δ\DeltamAP+0.0010+0.0010−0.0156-0.0156
Reviewer_o5fdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SIPA, a gray-box feature attack on cooperative end-to-end driving. It uses broadcast request maps to infer coarse route intent and guide a surrogate-based attack, while filtering updates to limit perception changes. Experiments on eight CARLA routes with a CoDriving-based system show reduced Driving Score and increased collisions, alongside small average perception changes and a CP-Guard rejection rate close to the clean false-alarm rate

Critical strengths
  1. The paper demonstrates a relevant gap between standard perception metrics and driving safety. Collisions increase from 0.94 to 9.04 per kilometer while average [email protected] changes by only −0.002 in the tested setting (Table 1). This is a useful finding for evaluating cooperative driving systems.

  2. The ablations support the contribution of request-map information. A size-matched random ROI reduces Driving Score by 11.20 points, compared with 19.91 for SIPA. Removing intent conditioning reduces the drop to 5.94 points (Tables 2–3).

Critical weaknesses
  1. The practical access assumptions need stronger justification. Sec. 3.1 grants the attacker access to the actual ego-side perception inputs and additional contextual tensors needed by the surrogate and perception filter. A shared perception implementation does not itself explain how a compromised collaborator obtains another vehicle’s runtime inputs. These assumptions are disclosed, but their practical availability remains unclear from the main text. This limits how directly the results establish a threat from a single compromised collaborator.

  2. The evidence for preserving perception is not sufficiently precise. Eq. 14 suppresses confidence increases at initially low-confidence locations, but does not directly constrain missed detections or localization changes. The aggregate mAP and precision results in Table 1 could therefore conceal a small number of perception errors with large safety consequences. Without examining detection changes around driving failures, it remains unclear whether SIPA selectively disrupts planning or introduces sparse but consequential perception errors.

Questions for authors
  1. Access clarification: Which ego-side inputs and contextual tensors are available to the compromised collaborator through the V2X protocol, and which require additional access? This would clarify the practical scope of the threat model.

  2. Validation of perception preservation: Do the small average perception changes also hold for nearby objects and individual failure events, including missed detections and localization errors? This would help establish whether the driving failures occur while relevant perception remains intact.

  3. Defense evaluation clarification: Were the Driving Score and collision results in Table 1 measured with CP-Guard actively removing rejected features, or was rejection evaluated separately? This distinction matters when interpreting the consequences for a defended driving system.

2 comments

We thank the reviewer for the careful comments. We respond to each point below.

W1 · Q1 — Practical access assumptions; which inputs come from the protocol vs. require additional access

Access in our threat model is partitioned by module, not by vehicle. The perception stack can be reproduced by the attacker under the assumption of a homogeneous deployment with publicly released checkpoints, whereas the victim planner's parameters, gradients, outputs, and route are not accessible.

More importantly, we re-evaluated the attack in a setting that uses none of the victim's ego-side inputs. We replace the ego BEV (X0) with zeros before perturbation injection; this applies to the entire perception forward pass, so even the detections and occupancy used by the perception filter are computed without X0 via an independent surrogate detector. We also do not use the victim's past 4-tick temporal bank, and we do not project the attacker's features into the victim frame to inject an estimated input. As shown below, under the original threat model SIPA attains ΔDS −19.91, and even in this most conservative setting that uses none of these inputs the attack is maintained with little change (ΔDS −14.90). Therefore, the concern of "how a compromised collaborator obtains the victim's runtime inputs" is largely resolved—the attack does not require those inputs.

To directly answer the reviewer's question on which inputs are available through the protocol versus require additional access: what is shared through the protocol is the request map, and SIPA uses only this as its attack signal. Runtime inputs that would require additional access—the ego inputs, past temporal state, etc.—are not used at all in the above setting. We will summarize the provenance of each input in the main text to make the scope of the threat model explicit.

MethodAttacker accessΔ\DeltaDSPrecision%
SIPAVictim runtime inputs−19.91-19.9196.19
SIPAProtocol-visible only−14.90-14.9092.30
SIPA w/o counter-intentProtocol-visible only−9.73-9.7394.56
SOMBRA*Victim runtime inputs−45.98-45.9828.00
SOMBRA*Protocol-visible only−35.29-35.2927.78
PGD*Victim runtime inputs−41.39-41.3915.91
PGD*Protocol-visible only−40.76-40.7619.87

W2 · Q2 — Precision of the perception-preservation evidence; nearby objects and individual failure events

We agree. Our filter (Eq. 14) only suppresses confidence increases at initially low-confidence locations and does not directly constrain missed detections or localization changes; hence the average mAP and precision in Table 1 could hide a small number of safety-critical errors.

To examine this, we aggregated the 10 frames immediately preceding each collision—i.e., around the driving-failure points—and measured perception changes for collision-relevant objects. As shown below, even in this window SIPA largely preserves perception (ΔmAP near zero, recall maintained), whereas over the same window SOMBRA* and PGD* incur large perception damage. Thus, even at the per-frame level right before failure, SIPA disrupts planning selectively while relevant perception is preserved, rather than introducing sparse but consequential perception errors.

SettingSegmentCollisionsFramesΔ\Delta[email protected]PrecisionRecallFP/frameFN/frameMismatchADE
No defenseFull drive41862+0.0016+0.001692.69%94.54%0.2270.1438.86%0.1031
No defensePre-collision 10f432−0.0036-0.003695.30%94.04%0.1880.2509.27%0.2438
τ=0.7\tau=0.7Full drive102005+0.0047+0.004792.95%92.09%0.2110.2648.34%0.1043
τ=0.7\tau=0.7Pre-collision 10f1068−0.0105-0.010592.98%94.93%0.3240.1475.59%0.1341

Q3 — Was Table 1 measured with CP-Guard actively removing rejected features, or evaluated separately?

To clarify: the Driving Score and collision results in Table 1 were obtained by evaluating rejection separately; they were not measured with the defense actually removing rejected features. We agree this distinction matters. The corresponding closed-loop experiment in the enforced condition—where CP-Guard actually removes rejected features—is currently in progress, and we will post the results (clean@enforced and SIPA@enforced Driving Score and collisions) as soon as they are available.

Update: we have completed the enforced closed-loop evaluation. The results are below.

MethodΔ\DeltaFP/fPrecision%Δ\DeltaDS
SIPA+0.08+0.0894.5%−15.70-15.70
SOMBRA*+15.99+15.9923.4%−33.66-33.66
PGD*+28.91+28.9114.8%−12.83-12.83

Enforcing the defense degrades the attacks unevenly, and SIPA is mitigated the least. SIPA's ΔDS stays nearly the same (−19.91 → −15.70), so meaningful driving harm persists under the defense, whereas the baselines are mitigated more (SOMBRA* −45.98 → −33.66, PGD* −41.39 → −12.83). This matches their heavy perception corruption: even under enforcement they retain far lower precision (23.4% / 14.8% vs. SIPA 94.5%) and much larger spurious detections (Δ\DeltaFP/f +15.99 / +28.91 vs. +0.08). Thus a perception-level defense mitigates the baselines more than SIPA. Notably, under the enforced defense SIPA (−15.70) even surpasses PGD* (−12.83), reversing their no-defense ranking.

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

While existing V2X cooperative driving attacks mainly targets perception, this paper proposes SIPA (Stealthy Intent Planning Attack) that exploits the V2X cooperative perception by transmitting a malicious BEV feature that only hurts the planning performance while keeping the perception performance less impacted. SIPA exploits a protocol-visible side channel (request maps generated from provisional waypoints) to condition a surrogate planner, infer a counter-intent steering direction, and optimize a control-aware planning objective without creating perception drift. Implementation on closed loop CARLA testbed shows the effectiveness of the proposed attack, reducing the Driving Score by 19.91 point.

Critical strengths
  • Interesting finding of the side channel information of cooperative perception that can be exploited by an adversary.
  • Design of planning-level attack without impacting the perception performance demonstrates an interesting insight about the disjoint nature of perception vs planning.
Critical weaknesses
  • The attack depends on the request map features that is not, as far as I am concerned, part of any standardized protocol. So, the attack is build on the other research initiatives in the line of the cooperative perception.
  • Attack success depends on the performance of the success of the surrogate planner but it remains unclear how such training would be feasible and practical in a real-world setting (given a limited compute, time, and resources).
  • Evaluation remains limited in some aspect. For instance, the stealthiness of SIPA is evaluated against CP-guard where CP-guard itself shows 19.2% rejection rate in benign scenario which ideally should be approximately 0. Hence, the justification of SIPA's stealthiness if weakened due to imperfection of the scale.
Questions for authors
  • What is the justification that CP-Guard is the right (and only one) metric to gauge the stealthiness of SIPA in the perception plane? What other methods can be used to justify the claim that perception and planning uses different underlying property of the same BEV feature and SIPA only exploit the planning related features?
  • The paper needs better justification and preliminary exploration (and results) to demonstrate the claim of side channel data leak in the BEV features about the planning.
  • Table 1 needs better justification on why the differences are not using the absolute value of the differenences such as | delta mAP |? Overall the evaluation needs to be more comprehensive. For example, it would be interesting to see the scalability of such as attack. How much time does it take to generate such an attack and whether it can comply with the V2X transmission rate.
4 comments

We thank the reviewer for the detailed and equation-grounded comments. We respond to each point below.

Q1 — Is CP-Guard the right (and only) metric for stealthiness? What else supports that perception and planning use different properties of the same BEV feature?

CP-Guard is not the only metric; we adopt it as a representative perception-level consensus defense. The claim that perception and planning rely on different properties of the same BEV feature—and that SIPA targets only the planning-related component—is directly supported by an ablation: removing the perception filter changes ΔmAP substantially (down to −0.258), while ΔDS in fact increases. This shows that the two tasks respond differently to the same feature.

W1 — The request map is not part of any standardized protocol

Request-map-based spatial feature requests are not specific to CoDriving; they are a widely used communication paradigm in cooperative perception. Many works adopt this approach—Where2comm, Select2Drive, CoSDH, Directed-CP, WhisperNet, Plan2comm, Why2comm—and in particular Select2Drive (Sec. IV-B, Eq. 22, Algorithm 1) exchanges with neighboring vehicles a Gaussian request map generated from previously planned waypoints.

Moreover, sharing future routes or driving intent itself has precedents in official V2X standards. For example, SAE J2735 defines DF_PathPrediction, SAE J3186 specifies maneuver sharing, SAE J3216 classifies intent sharing, and ETSI TR 103 578 describes planned-trajectory exchange.

We do not, however, claim that the specific implementation of a continuous BEV request map is itself standardized. Our point is that this is an actively studied design with analogous precedents in standards, and that proactively exposing its security vulnerability ahead of standardization is a contribution of this work.

W2 — Feasibility of training the surrogate planner

We will specify the surrogate planner's training procedure in the main text (currently Appendix A.7). The surrogate is trained from random initialization on Towns 1–4 and 6, with Town05 held out, and it never loads the victim planner's checkpoint at any stage. This is a one-time offline cost performed with publicly available simulators and data, so it is practical regardless of real-time compute resources.

Q2 — Stronger justification/exploration for side-channel leakage of planning information in BEV features

To support the claim that the request map carries planning-related information, we directly measured the accuracy of route-intent recovery from the request map. Rather than arguing this qualitatively, in the final setting (κ=6) we compared the recovered path Ψ(R) with the actual planned waypoints in the same frame, over 8 Town05 routes × 3 repetitions (2,508 attacked frames in total). As shown below, the recovery error is small relative to the victim path length (median ADE about 3% of the path length), does not accumulate along the trajectory, and the directional sign agrees in a high fraction of frames (87.2%). Recovery succeeds on 75.0% of frames; the remaining cases are near-stationary or straight segments where the waypoints collapse into a single component—i.e., cases with no direction to recover rather than failures of recovery (handled by the fallback in Appendix A.3). This directly shows that the spatial structure of the request map carries coarse route intent.

QuantityMedianMean
ADE0.226 m0.330 m
FDE ∥q^−wT∥\|\hat{q}-w_T\|0.259 m0.497 m
Path-length ratio1.0071.047
Recovery success rate75.0% (1,880 / 2,508)
Counter-intent sign agreement87.2% (1,250 / 1,433)

W3 — CP-Guard's benign false-alarm rate (19.2%) is too high, weakening the stealthiness argument

The benign false-alarm rate is far from the ideal of 0, and at this scale the baseline for stealthiness comparison can be loose. To clarify: the 19.2% the reviewer refers to corresponds to the original setting (τ=0.93); at that setting the clean rejection rate is 19.2% and SIPA's is 18.4%, i.e., essentially identical. To address this, we lowered τ to stricter values so that the clean rejection rate decreases, and re-evaluated all attacks. The table below reports rejection rates (%) over τ=0.70–0.95.

τ\tauCleanSIPASOMBRASOMBRA*PGD*
0.706.97.59.029.753.3
0.757.48.19.034.255.2
0.808.89.610.340.659.8
0.8512.912.513.147.963.1
0.9015.715.316.056.167.5
0.9319.218.417.861.471.2
0.9522.925.120.566.773.2

Across the entire range, SIPA's rejection rate stays close to clean (e.g., at τ=0.93, clean 19.2 vs. SIPA 18.4; at τ=0.80, 8.8 vs. 9.6), whereas SOMBRA* and PGD* are several times higher (29.7–66.7 and 53.3–73.2, respectively). Thus, regardless of the threshold choice, SIPA's conclusion—disrupting planning while preserving perception—holds, and it does not arise from a loose scale of the measurement tool.

Q3 — Why not use the absolute value |ΔmAP| in Table 1? Scalability, generation time, and V2X transmission rate

We do not use |ΔmAP| because the sign carries meaning: it distinguishes whether a perturbation degrades perception (negative) or preserves it (near zero), and taking the absolute value would conflate "almost no change" with "degradation." As requested, we will state the sign convention explicitly and also include a |ΔmAP| column.

Regarding scalability, we answer along two axes. (1) Computation/time: the per-frame attack generation time is about 17,053 ms under the default setting (K=60, ΔDS −19.91). Reducing the number of iterations to K=10 shortens it to about 2,850 ms, at which point ΔDS is −15.24—a slight trade-off, but still substantial. Compared with a typical V2X transmission interval (about 100 ms), we acknowledge that the current setting is heavy for real-time use. That said, this work aims to first present and demonstrate this attack surface, and the fact that the attack remains effective even with far fewer iterations suggests the possibility of online/real-time deployment through further optimization (e.g., warm-starting), which we leave to future work. (2) Model generalization: the attack is not specific to CoDriving's architecture; we confirmed that the same trend holds on another fusion model, V2X-ViT. We acknowledge the time constraint for the online setting, but it does not weaken this work's central claim—the existence of this attack surface.

MetricCoDriving (SIPA)V2X-ViT (SIPA)
Clean →\to attacked DS83.23 →\to 68.0246.86 →\to 35.67
Δ\DeltaDS−15.21-15.21−11.19-11.19
Δ\DeltamAP+0.0010+0.0010−0.0156-0.0156
Reviewer_NMAqOpenReview ↗

Assessment of Authors’ Rebuttal: Addressed and Remaining Concerns

Thank you for the detailed rebuttal and additional experimental results. My assessment of the key concerns is as follows:

Request Map Standardization (Largely Addressed). The additional references establish that request-map-based communication is actively explored in cooperative perception research, although the specific implementation is not standardized.

Surrogate Planner Feasibility (Partially Addressed). The clarification regarding offline training and independence from the victim's model is helpful. However, the training cost, data requirements, and practical feasibility under realistic adversarial constraints remain unclear.

Stealthiness Evaluation (Partially Addressed). The additional CP-Guard threshold analysis strengthens the results. However, evaluating only CP-Guard is insufficient to establish broader stealthiness.

Attack Scalability and Real-Time Feasibility (Not Adequately Addressed). Although reducing the optimization iterations improves efficiency, the reported 2.85-second attack generation time remains significantly higher than the approximately 0.10 second V2X transmission interval. The authors acknowledge this limitation but provide no experimental evidence of real-time feasibility.

Overall Assessment: The rebuttal provides meaningful additional evidence and addresses several concerns. However, the limited stealthiness evaluation and lack of demonstrated real-time attack feasibility remain significant weaknesses.

We thank the reviewer. We agree stealth should not rest on CP-Guard alone, and in particular that CP-Guard’s 19.2% benign false-alarm rate is an imperfect yardstick (NMAq). We address both concerns by evaluating SIPA against a second consensus defense, ROBOSAC (Li et al., 2023, “Among Us” — already in our references), which (i) uses a different verification mechanism and (ii) has a benign false-alarm rate close to the ideal 0%. ROBOSAC does not score a rasterized consistency map against a calibrated threshold. It repeatedly samples subsets of collaborators and accepts a subset only if the fused detection is geometrically consistent (IoU/Jaccard) with the ego-only detection, otherwise falling back to ego-only. It thus verifies collaborators through an independent signal with a much lower benign rejection rate. We run ROBOSAC under the exact Table-1 setting (same eight Town05 routes, single compromised RSU, same independently-trained surrogate, ε = 6, K = 60) and report the collaborator rejection rate (Detection), frame-pooled as in the paper; for Clean this is the benign false-alarm rate.

MethodΔ[email protected] (rel. to clean)ROBOSAC Detection ↓
Clean—0.6% (benign false-alarm)
SIPA (Ours)−0.002 (−0.6%)0.6%
SOMBRA*−0.282 (−85%)56.4%

Parenthetical = drop relative to the clean [email protected] ≈ 0.34 (absolute: SIPA 0.340→0.338; SOMBRA* 0.332→0.054). Δ[email protected] from Table 1. Detection is frame-pooled over verified frames: Clean 13/2105, SIPA 7/1131, SOMBRA* 317/562.

Two points. (1) The stealth conclusion does not depend on CP-Guard’s inflated benign scale. ROBOSAC’s benign false-alarm rate is 0.6% — essentially the ≈0% baseline the reviewer expected — yet SIPA is rejected at exactly that rate (0.6%), i.e., statistically indistinguishable from an honest collaborator, while the perception-damaging SOMBRA* is rejected on 56.4% of verified frames (≈90× higher). (2) The ordering again tracks perception damage, not driving damage (SIPA ΔmAP = −0.002 vs SOMBRA* −0.282). This matches the paper’s thesis: a perception-level consensus check — whether it rasterizes boxes (CP-Guard) or samples subsets (ROBOSAC) — cannot flag a planning attack that preserves cooperative perception.

Reviewer_CfkHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies perception-stealthy planning attacks against V2X end-to-end autonomous driving and proposes SIPA. The attacker controls a single collaborator and perturbs its shared BEV features to disrupt the victim vehicle’s planning while limiting changes to detection outputs. The method recovers coarse route intent from the spatial structure of planning request maps and uses it to condition an independently trained surrogate planner, select perturbation regions, and construct a counter-intent control objective. It further applies perception-aware gradient filtering to suppress spurious detection responses.

Critical strengths
  1. Reinterpreting request maps, originally designed for communication efficiency, as a side channel for route intent is the paper’s most distinctive contribution. The authors further exploit this signal to guide planning attacks, extending the scope beyond detection corruption in conventional cooperative perception attacks. This finding provides a concrete basis for examining the security of planning-aware communication protocols.
  2. The experiments report Driving Score and collision rates, capturing the attack’s consequences for driving behavior. The same-frame clean counterfactual obtained by removing the perturbation, together with a random-ROI baseline matched in perturbation area on each frame, helps reduce confounding from scene differences and perturbation size.
Critical weaknesses
  1. The evaluation does not sufficiently cover the conditions required for the attack to work. Appendix A.7 states that attack optimization uses the actual ego-side perception inputs, as well as temporal and contextual tensors supplied by the evaluation pipeline. However, access to a shared perception model does not automatically mean that an ordinary malicious collaborator can obtain these runtime inputs. The paper does not specify how each input is communicated, so it is unclear how much of the attack’s effectiveness would remain if it relied only on protocol-visible information. Appendix A.4 also states that the experiments use global target weight = 0, which allows the final predicted trajectory segment to directly affect steering. At the default value of 1, this direct steering pathway is removed. Since the control-aware objective is a central part of the method, the lack of results under the default configuration limits the evidence for broader claims about the safety of driving systems.
  2. The current metrics do not sufficiently establish perception stealth or defense evasion. The perception filter only suppresses confidence increases at locations that had low confidence in the clean prediction; it does not directly constrain missed detections, bounding-box shifts, or confidence decreases for existing objects. A similar average mAP does not rule out significant perception errors on a small number of collision-critical frames. The authors should report same-frame matching and recall for relevant objects, together with a frame-by-frame analysis before collisions, to establish whether the driving disruption occurs while detections remain substantially unchanged. For the defense evaluation, the authors replace CP-Guard’s segmentation representation with a rasterized bounding-box representation and recalibrate the normalization and threshold. The single-collaborator setting also reduces its subset-verification mechanism to a single verification. The rejection rate at one threshold is insufficient to characterize detection performance across operating points, and the paper does not clearly report closed-loop driving results when rejection is actually enforced.
  3. The mathematical notation and definitions of key variables are incomplete. Several symbols are not explained when first introduced, including(λ), (v_max), (τ_p), and the averaging notation in Equations (12)–(14). Some of these are later defined in the appendix, but the Gaussian scale parameter (ς) in Equations (24)–(25) remains unclear. The units and transformations between request-map coordinates, recovered-path coordinates, and planner-input coordinates are also unspecified. These gaps concern core steps in intent recovery and direction estimation, making the method harder to understand and reproduce. The authors should provide a consistent notation table and clearly specify the coordinate axes, units, and transformations.
Questions for authors
  1. Please specify each ego feature, historical input, and contextual tensor used by the attack, and explain how each is obtained through normal protocol messages. How effective is the attack when using only protocol-visible information or attacker-estimated inputs?
  2. Why do the experiments use global target weight = 0? Under the default value of 1, what are the clean driving performance, attacked Driving Score, and collision rate? Is the observed degradation primarily due to steering deviations or deceleration?
  3. Do any key objects become missed or shift in the detections before a collision? Does CP-Guard’s rejection actually trigger feature removal or degradation? How much driving harm remains when the defense is enforced?
  4. Does SIPA retain an advantage when the same perception filter is added to PGD* or SOMBRA*, or when those baselines are adjusted to achieve comparable perception drift? Can the authors report uncertainty across routes and attack runtime per frame?
4 comments

We thank the reviewer for the detailed, equation-grounded comments. We respond to each point below.

Q1 · W1 — Inputs used by the attack and their provenance; effectiveness under protocol-visible / estimated inputs only

Access in our threat model is partitioned by module, not by vehicle. The perception stack is white-box under the assumption of a homogeneous deployment with publicly released CoDriving checkpoints (so the attacker can reproduce it on the shared scene), whereas the victim planner's parameters, gradients, predictions, and route remain black-box throughout.

To address the reviewer's concern more directly, we re-evaluated the attack in a setting that uses none of the victim's ego-side inputs. Specifically, (a) we replace the ego BEV (X0) with zeros at the scatter stage, before perturbation injection; this applies to the entire perception forward pass, so even the detections and occupancy used by the perception filter are computed without X0 via an independent surrogate detector. In addition, (b) we do not use the victim's past 4-tick temporal bank, and we do not project the attacker's features into the victim frame to inject an estimated (warped) input. This is a stricter setting than either of the two cases the reviewer mentions ("protocol-visible information only" or "attacker-estimated inputs only").

As shown below, under the original threat model SIPA attains ΔDS −19.91, and even in this most conservative setting that uses none of these inputs the attack is maintained with little change (ΔDS −14.90). In contrast, removing the recovered route intent under the same ego-free condition (w/o CI) drops the effect substantially to −9.73. This indicates that the attack's effectiveness originates not from access to the victim's ego inputs but from the protocol-visible request map. SIPA also maintains high detection precision in both settings (96.19 / 92.30), whereas SOMBRA* and PGD* show low precision.

MethodAttacker accessΔ\DeltaDSPrecision%
SIPAVictim runtime inputs−19.91-19.9196.19
SIPAProtocol-visible only−14.90-14.9092.30
SIPA w/o counter-intentProtocol-visible only−9.73-9.7394.56
SOMBRA*Victim runtime inputs−45.98-45.9828.00
SOMBRA*Protocol-visible only−35.29-35.2927.78
PGD*Victim runtime inputs−41.39-41.3915.91
PGD*Protocol-visible only−40.76-40.7619.87

We will summarize in the main text which inputs are provided through the protocol and which require additional access. The concern about global target weight = 0, also raised in W1, is addressed in Q2 below.

Q2 — Why global target weight = 0; performance under the default (=1); steering vs. deceleration

weight = 0 is a deliberate choice for correctly measuring the planning attack, which we will clarify in the main text. Under global target weight = 1 (the stock setting), the controller steers directly toward the ground-truth global target, so the predicted trajectory has little influence on steering. Evaluating a planning attack under weight = 1 therefore structurally masks its effect, because steering bypasses the (attacked) prediction and follows privileged ground-truth route information. We used weight = 0 to remove this shortcut so that the predicted trajectory actually drives steering.

W2 · Q3 — Perception stealth beyond average metrics; detection changes before collisions; defense operating points

We agree with this point. Our filter only suppresses confidence increases at locations that were low-confidence in the clean prediction, and does not directly constrain missed detections or localization changes; hence a similar average mAP does not rule out consequential perception errors on a small number of collision-critical frames.

To examine this, we aggregated only the 10 frames immediately preceding each collision and measured perception changes for collision-relevant objects. As shown below, even in this window SIPA largely preserves perception (ΔmAP near zero, recall maintained), whereas over the same window SOMBRA* and PGD* incur large perception damage. This supports that SIPA's driving failures arise from planning disruption rather than perception errors.

SettingSegmentCollisionsFramesΔ\Delta[email protected]PrecisionRecallFP/frameFN/frameMismatchADE
No defenseFull drive41862+0.0016+0.001692.69%94.54%0.2270.1438.86%0.1031
No defensePre-collision 10f432−0.0036-0.003695.30%94.04%0.1880.2509.27%0.2438
τ=0.7\tau=0.7Full drive102005+0.0047+0.004792.95%92.09%0.2110.2648.34%0.1043
τ=0.7\tau=0.7Pre-collision 10f1068−0.0105-0.010592.98%94.93%0.3240.1475.59%0.1341

We also agree that a rejection rate at a single threshold is insufficient to characterize detection performance across operating points. First, to clarify the benign rejection rate: at the original setting (τ=0.93), the clean rejection rate is 19.2% and SIPA's is 18.4%, i.e., essentially identical. The table below reports rejection rates (%) over τ=0.70–0.95.

τ\tauCleanSIPASOMBRASOMBRA*PGD*
0.706.97.59.029.753.3
0.757.48.19.034.255.2
0.808.89.610.340.659.8
0.8512.912.513.147.963.1
0.9015.715.316.056.167.5
0.9319.218.417.861.471.2
0.9522.925.120.566.773.2

Across the entire range, SIPA's rejection rate stays close to clean (at τ=0.93, clean 19.2 vs. SIPA 18.4; at τ=0.80, 8.8 vs. 9.6), whereas SOMBRA* and PGD* are several times higher (29.7–66.7 and 53.3–73.2, respectively). Thus the stealthiness conclusion holds regardless of the threshold choice.

Regarding the representation, we rasterize bounding boxes onto a BEV occupancy representation only to match our detector's output, and we calibrate the threshold on clean frames. We also note that with a single collaborator, subset verification reduces to a single check, and we interpret these results as an evaluation of the per-collaborator verifier. As for whether CP-Guard's rejection actually triggers feature removal/degradation and how much driving harm remains under an enforced defense, the corresponding closed-loop experiment—with rejected features actually removed—is currently in progress, and we will post the results as soon as they are available.

Q4 — Advantage when the same filter is added to baselines; uncertainty across routes; per-frame runtime

We added PGD*+filter and SOMBRA*+filter, which apply the identical perception filter, to check whether SIPA retains its advantage when perception drift is matched. As shown below, even under the same filter, SIPA achieves the largest Driving Score degradation (ΔDS) and the lowest matching mismatch rate.

Method (+ filter)Δ\DeltaDSΔ\DeltamAP(%)Mismatch%
SIPA−14.90-14.90−2.2-2.29.66
SOMBRA*−13.37-13.37−10.2-10.210.02
PGD*−5.22-5.22−6.4-6.414.49

We will also report all headline numbers as route-wise mean ± std. The per-frame attack generation time is about 17,053 ms under the default setting (K=60, ΔDS −19.91). Reducing the number of iterations to K=10 shortens it to about 2,850 ms, at which point ΔDS is −15.24—a slight trade-off in attack strength, but still substantial. Compared with a typical V2X transmission interval (about 100 ms), we acknowledge that the current setting is heavy for real-time use. That said, this work aims to first present and demonstrate this perception-preserving, planning-disrupting attack surface; the observation that the attack remains effective even with far fewer iterations suggests the possibility of real-time deployment through further optimization (e.g., warm-starting), which we leave to future work.

W3 — Incomplete notation and definitions of key variables

We will add a consistent notation table that defines each symbol at its first appearance—λ\lambda, vmax⁡v_{\max}, τp\tau_p, the spatial-average notation ⟨⋅⟩\langle\cdot\rangle in Eqs. (12)–(14), and the Gaussian scale parameter ς\varsigma in Eqs. (24)–(25). We will also specify the axes, units, and transformations among the request-map, recovered-path, and planner-input coordinates. In addition, we will make explicit that Eq. (24) is a geometric interpretation rather than an exact inverse transform.

SymbolMeaningFirst appearsValue / Note
λ\lambdaWeight of the speed term in the control-aware objectiveEq. 12 (= Eq. 39)λ=1\lambda = 1
vmax⁡v_{\max}Normalization / upper bound on desired speedEq. 12, 36–37vmax⁡=8v_{\max} = 8
⟨⋅⟩\langle\cdot\rangleSpatial meanEq. 14Same as mean[⋅]\mathrm{mean}[\cdot] in Eq. 43
τp\tau_pPerception-filter confidence thresholdEq. 14τp=0.2\tau_p = 0.2
ς\varsigmaGaussian width of request-map renderingEq. 24–25Geometric interpretation, not an exact inverse
a(W)a(W)Steering-related quantity fsteer(W)f_{\text{steer}}(W)Eq. 10, 32–34Depends on the final segment wT−wT−1w_T - w_{T-1}
v(W)v(W)Desired-speed-related quantity fspeed(W)f_{\text{speed}}(W)Eq. 10, 35–37(∥w4∥−∥w1∥)/3⋅vmax⁡(\lVert w_4\rVert - \lVert w_1\rVert)/3 \cdot v_{\max}
ssCounter-intent steering signEq. 8, 30s=−sign(φR)s = -\mathrm{sign}(\varphi_R)
q^\hat{q}Recovered route goalEq. 6, 28=wTf= w^f_T (recovered endpoint); pixel-to-planner mapping required

Update: we have completed the enforced closed-loop evaluation. The results are below.

MethodΔ\DeltaFP/fPrecision%Δ\DeltaDS
SIPA+0.08+0.0894.5%−15.70-15.70
SOMBRA*+15.99+15.9923.4%−33.66-33.66
PGD*+28.91+28.9114.8%−12.83-12.83

Enforcing the defense degrades the attacks unevenly, and SIPA is mitigated the least. SIPA's ΔDS stays nearly the same (−19.91 → −15.70), so meaningful driving harm persists under the defense, whereas the baselines are mitigated more (SOMBRA* −45.98 → −33.66, PGD* −41.39 → −12.83). This matches their heavy perception corruption: even under enforcement they retain far lower precision (23.4% / 14.8%) and much larger spurious detections (Δ\DeltaFP/f +15.99 / +28.91) than SIPA (94.5%, +0.08). In short, a perception-level defense mitigates the baselines more than SIPA. Notably, under the enforced defense SIPA (−15.70) even surpasses PGD* (−12.83), reversing their no-defense ranking.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.