acceptodds
Under review as a conference paper at ICLR 2027

Don't Let One Lie Survive A Hundred Truths: A Selective Bayesian Trust Estimator for Collaborative Perception

Abstract

Collaborative perception (CP) enables connected vehicles to see beyond their own sensors but makes them dependent on messages they cannot independently verify. A compromised collaborator can surgically conceal a single safety-critical object or inject a non-existing one while correctly reporting many others. Existing Bayesian trust mechanisms pool agreement across objects, which, while effective against blatant untargeted attacks, either incurs high false-positive rates (FPR), or allows unrelated correct reports to dilute persistent attack evidence for stealthy single-object attackers. To address this problem, we propose SABER, a selective two-tier Bayesian trust estimator. The first tier maintains broad agent and object trust, preserving the ability to downweight benign but low-quality contributors. Cumulative-sum screening selects agent–object pairs with persistent omissions or unsupported reports for focused Bayesian assessment. The second tier checks these pairs against other agents’ evidence and maintains a separate, reference-weighted Beta state for each. The lowest pair score constrains agent trust, preventing unrelated reports from diluting a targeted attack. We establish sufficient conditions for stronger attacker-side trust reductions with bounded additional benign false alarms at fixed thresholds. Compared with state-of-the-art CP defenses, SABER improves attack detection while reducing benign FPRs. On OPV2V, SABER improves defense ROC-AUC over MATE by up to 0.427 in late fusion and 0.337 in intermediate fusion. Against advanced intermediate-fusion data fabrication attacks, it increases detection rates over ROBOSAC and LUCIA by up to 96.40 and 67.07 percentage points, respectively, while reducing FPRs.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Cyber-Physical System Security, Collaborative Perception, Bayes' Theorem

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6nNsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a fairly specific problem in collaborative perception: if an attacker repeatedly omits or fabricates one critical object while reporting other objects correctly, the vehicle may still receive a high trust score when all its reports are pooled. The authors propose SABER, which retains an overall trust estimate but also uses CUSUM to identify agent–track pairs with persistent anomalies. It then uses observations from other vehicles to update a separate trust score for each selected pair. A low score for any pair under assessment can lower the vehicle’s final trust score. To me, the central idea is to prevent a large number of correct reports from masking a sustained attack on a single object. The authors evaluate object removal and spoofing attacks on OPV2V under both late and intermediate fusion. Based on the reported results, SABER improves the trade-off between detection and false alarms, with particularly strong results against sustained spoofing attacks.

Critical strengths
  1. The theoretical analysis is very thorough. I consider this the most valuable part of the paper. The authors not only propose a mechanism for keeping separate evidence for each agent–track pair, but also analyze cross-object evidence dilution, reductions in trust scores, and false-alarm control. They state the conditions under which their conclusions hold.

  2. The paper targets a specific problem, and the method is relatively easy to interpret. Compared with methods that check whether an agent’s perception agrees with a reference vehicle’s, or rely on learned feature differences associated with malicious behavior, this paper focuses on an attacker that reports most objects correctly while persistently manipulating one critical object. I think such an attack could indeed be masked by an aggregate trust score. Keeping separate evidence for each agent–track pair is a clear and meaningful way to preserve these localized inconsistencies.

Critical weaknesses

The method relies on restrictive assumptions. It considers only one attacker, although colluding attackers could manipulate the peer reference. Even with one attacker, a genuine detection may be indistinguishable from a fabricated report if only one vehicle can observe the object. Since some test cases require strong coverage by benign vehicles, I cannot tell from the current results how reliable SABER would be under weaker coverage or collusion.

Questions for authors
  1. The threat model considers only one attacker. How does SABER perform with different proportions of malicious vehicles, particularly when several attackers collude or outnumber benign vehicles?

  2. How exactly is the threshold η\eta chosen? Would the current setting transfer to other scenarios?

  3. I am curious about the cold-start case: if an attacker strikes before much historical evidence has accumulated, would the Tier 1 trust estimate and subsequent detection still be effective? Experiments with attacks starting at frame 0 and different warm-up lengths would help answer this.

  4. Latency seems particularly important for an agent–track-pair-based defense in autonomous driving. Yet the experiments do not report the number of frames needed to raise an alarm, or the per-frame computation time and memory overhead. It is also unclear whether the final fused perception improves with the defense (for example, in [email protected] or [email protected]) or how much performance is lost when a benign vehicle is mistakenly flagged.

  5. Consider two common cases: an attacker fabricates an object, and a collaborator detects a real object for the first time but is the only vehicle able to see it. If the other vehicles cannot observe that location and have no relevant history, the system appears to receive the same information in both cases. Can SABER distinguish them? If not, the paper should clarify the peer-observation conditions it requires and report results when they are not met.

  6. The experiments rely mainly on a single backbone, which makes it difficult to assess generalization. Would SABER remain effective with a newer fusion model and backbone, and would its parameters need recalibration? An evaluation with a structurally different model would be useful, though this is not a deciding factor in my recommendation.

  7. A couple of minor presentation issues remain. One piece of text in Figure 1 ends with a period while the others do not. Also, Equation (1) uses �1��1� for TPR but �2��2� for FPR; the intended distinction is unclear.

Reviewer_L4D2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SABER, a two-tier Bayesian trust estimator for collaborative perception. It uses CUSUM screening, peer-based object-specific trust updates, and minimum aggregation to detect persistent targeted removal and spoofing attacks. Experiments on OPV2V report improved detection–false-positive trade-offs against several existing defenses.

Critical strengths
  1. Object-specific trust estimation directly addresses evidence dilution in pooled Bayesian models, a meaningful failure mode for sparse targeted attacks.

  2. The evaluation covers multiple attack types, fusion settings, and attack durations, with component ablations and ROC comparisons supporting the empirical benefits.

Critical weaknesses
  1. Appendix C.3 selects persistent, observable targets with substantial benign-agent coverage, closely matching SABER's assumptions. Generalization to less favorable but realistic conditions is not established.

  2. Theoretical results largely depend on assumed evidence separation, while the experimental evaluation does not demonstrate robust end-to-end safety improvements or isolate the practical consequences of delayed detection.

Questions for authors

(1) Appendix C.3 requires spoofing positions to be visible to at least two benign agents. How much does detection deteriorate with only one corroborating observer?

(2) Section 3 allows falsified FOV declarations but excludes denial-of-collaboration. Can an attacker evade omission detection simply by manipulating visibility declarations?

(3) Table 6 shows late-fusion spoofing TPR rising from 1.24% at five frames to 100% at ten. What explains this discontinuity, and how many attacks are detected before causing harm?

(4) Proposition 1 fixes reference and screening histories, while Theorems 5–6 assume favorable evidence bounds. What nontrivial guarantees remain when these quantities are jointly affected by attacks?

(5) Appendix C.4 generates attacks against an undefended centralized model. Do the reported gains persist under adaptive white-box attacks targeting SABER's screening and trust updates?

(6) Table 1 reports agent-flagging rates rather than protected perception quality. Does SABER improve downstream detection AP or collision outcomes after rejecting suspected collaborators?

Reviewer_k9o2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SABER, a two-tier trust estimator for collaborative perception under persistent, sparse object-removal and spoofing attacks. Experiments on OPV2V evaluate attacker detection across late and intermediate fusion settings.

Critical strengths
  • The motivation is clear. Cross-object evidence dilution is a meaningful limitation of aggregated trust estimation.
  • Excluding the assessed sender from both current observations and historical reference updates addresses self-reinforcement.
Critical weaknesses
  • As the vehicle's perception might be compromised, a robust controller under state uncertainty is not considered. Some threat could be handled by the controller.
  • Please add comparison and literature review for other anomaly detection algorithms.
  • There is no hardware experiment to validate the effectiveness of the method. Evaluation on OPV2V dataset is limited.
Questions for authors
  • How do you solve the deadlock issue among multiple vehicles? When each vehicle could receive some threats, they need to wait for each other to make decisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.