acceptodds
Under review as a conference paper at ICLR 2027

FedPurifier: Proactive Distributed Experience Purification for Resilient Connected Autonomous Vehicle Platoons

Abstract

Multi-agent reinforcement learning and graph attention networks are widely applied to cooperative control in connected autonomous vehicle platoons. However, anomalous experiences from local sensor faults or backdoor attacks propagate to neighbors via inter-vehicle communication and infiltrate the global policy through parameter sharing, triggering cascading platoon collisions. In conventional solutions, pre-filtering corrupted data struggles to detect stealthy anomalies, rigid behavioral constraints induce coordination deadlocks, and retraining from scratch incurs prohibitive computational and time costs. To address these challenges, we propose FedPurifier, an experience purification framework tailored for connected vehicle platoons. While maintaining fleet-wide parameter sharing, FedPurifier introduces orthogonal communication signatures to pinpoint responsible vehicles upon collision, performs contrastive unlearning across dual action and communication spaces to eliminate anomalous experiences, and preserves nominal driving capabilities via parameter anchoring. Across four MetaDrive benchmarks with 100 independent random seeds, FedPurifier on average reduces the backdoor attack success rate to 6.6% and restores the safety rate to 61.9%. The purified policy matches complete retraining performance while accelerating computation by two to three orders of magnitude.

Primary area:
reinforcement learning
Keywords:
Multi-Agent Reinforcement Learning, Machine Unlearning, Connected Autonomous Vehicles, Cooperative Control, Graph Attention Networks, Robustness and Safety

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jUoHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose FedPurifier, an experience purification framework tailored for connected vehicle platoons. While maintaining fleet-wide parameter sharing, FedPurifier introduces orthogonal communication signatures to pinpoint responsible vehicles upon collision, performs contrastive unlearning across dual action and communication spaces to eliminate anomalous experiences, and preserves nominal driving capabilities via parameter anchoring. Experimental results demonstrate the effectiveness of the method.

Critical strengths
  1. The authors conducted comprehensive experiments to analyze the effectiveness of the proposed method.
  2. The proposed framework considers both action-level and communication-level effects of anomalous experiences.
Critical weaknesses
  1. The paper does not explicitly explain why the signal response from a vehicle involved in a collision is significantly stronger than that of a normal vehicle. Theoretical analysis assumes the existence of this difference rather than demonstrating it based on vehicle behavior or communication characteristics.
  2. The threat model and attack assumptions are not sufficiently clear.
  3. This paper does not fully and directly evaluate performance under varying numbers of attacked vehicles, wireless channel conditions, and packet loss rates.
Questions for authors
  1. How does the orthogonal communication signature mechanism distinguish a vehicle that causes a collision from a benign vehicle participating in the same communication event?
  2. How is the stated perturbation bound satisfied by the backdoor trigger specified in Section 2, and what attacker capabilities are assumed after the purification process?
7 comments
AuthorsOpenReview ↗

[Response to Reviewer jUoH — Part 1/4]

Dear Reviewer jUoH,

Thank you for acknowledging the breadth of our experiments and our joint action-communication perspective. We address your comments point by point below:

Response to Q1 (W1 & Q1): Combining vehicle behavior and wireless communication properties, how do orthogonal signatures distinguish collision-causing vehicles from innocent vehicles in the same event?

Our orthogonal communication basis separates concurrent wireless signals at the physical layer without cross-talk, while collision responsibility is determined by each vehicle's kinematic prediction deviation during the 0.5 s pre-collision window. This decouples sender identity from crash causation in the same communication event and derives the response gap directly from vehicle behavior and channel physics rather than assumption. Behaviorally, each vehicle scales its signature weight by its one-step state-prediction residual projected onto the policy-sensitive direction: benign peers and braking victims follow nominal dynamics with tiny errors, whereas the initiator commands abrupt acceleration while hiding warnings. Physically, pairing 128-dimensional orthogonal carriers with pilot-based power pre-equalization eliminates inter-vehicle interference and near-far path-loss aliasing, so matched filtering at the RSU yields a background-normalized score uku_k that faithfully reflects behavioral anomaly and clears Theorem 1's conditional detection margin.

Table R3-1 empirically validates this separation in concurrent V2X events: the collision initiator yields 3.66× to 6.86× larger prediction residual norms than braking victims and benign neighbors, achieving 99.4% attribution accuracy with false-positive rates at or below 0.4%, while omitting attribution via w/o Attr rebounds ASR from 5.1%/5.4% to 18.5%/22.8%.

Table R3-1: Three-role attribution separation and unattributed ablation across 100 traffic seeds (bold highlights initiator metrics, new w/o Attr, and full FedPurifier).

Vehicle Role / VariantResidual NormNorm. ScoreAcc. / FPR (%)RB Safe / ASR (%)IS Safe / ASR (%)
Collision Initiator (Responsible)3.84 ± 0.290.91 ± 0.0699.4 ± 0.4 (Acc.)––
Passive Collision Victim (Braking)1.05 ± 0.180.23 ± 0.050.4 ± 0.3 (FPR)––
Benign Communicating Neighbor0.56 ± 0.120.13 ± 0.040.2 ± 0.1 (FPR)––
w/o Attr (Table 2: No Attribution)–––68.4 / 18.558.6 / 22.8
FedPurifier (Full)–––74.1 / 5.163.9 / 5.4

In conclusion, Table R3-1 and our decoupled physical-kinematic design confirm accurate attribution without penalizing innocent co-located peers.

Corresponding Manuscript Revisions:

  • Sec. 3.1 Eqs. (2)–(5), Theorem 1 & Trajectory Paragraph (pp. 4–5): Paired pilot-based power pre-equalization at the physical layer with one-step innovation-residual modulation, normalized score uku_k, conditional Theorem 1, and a dedicated pre-collision three-role trajectory differentiation paragraph.
  • Sec. 4.4 Table 2, Sec. 4.5 Table 3 & App. A.1, B.1 (pp. 8–9, 13, 15–16): Added Sec. 4.5 Table 3 validating separation across initiators (3.84, 99.4% accuracy), braking victims (1.05, 0.4% FPR), and benign peers (0.56, 0.2% FPR), added w/o Attr in Table 2, and expanded App. A.1 and App. B.1.
AuthorsOpenReview ↗

[Response to Reviewer jUoH — Part 2/4]

Response to Q2 (W2 part a & Q2 part a): How does the backdoor trigger in Section 2 satisfy the stated 0.05 perturbation bound?

The -0.15 m relative distance offset (Δd) and +0.20 m/s relative velocity offset (Δv) in Section 2 are defined in raw physical sensor units, whereas the 0.05 infinity-norm bound governs standardized neural policy inputs; to remove any notation ambiguity, revised Sec. 2 and Appendix A.2 now explicitly document the exact coordinate normalization procedure. Prior to policy ingestion, raw radar observations are divided componentwise by their nominal dynamic sensing scales sigma_norm = [5.0 m, 4.0 m/s], yielding normalized shifts -0.15 m / 5.0 m = -0.03 (for Δd) and +0.20 m/s / 4.0 m/s = +0.05 (for Δv). The resulting dimensionless trigger vector delta_trigger = [-0.03, +0.05] has exact infinity norm max(|-0.03|, |+0.05|) = 0.05 <= 0.05.

Because this 0.05 shift stays hidden within normal radar measurement noise variance and perturbs local parameter update norms by less than 1%, Byzantine weight-norm filters (Sec. 4.2 and Appendix A.3) fail to block it: Krum retains 48.6% (RB) and 81.4% (IS) ASR while Trimmed-Mean retains 46.9% and 79.2%. Triggered post-hoc from closed-loop collision residuals instead, FedPurifier drives ASR down to 5.1% and 5.4%.

In conclusion, physical-to-normalized radar scaling reconciles [-0.15 m, +0.20 m/s] with the 0.05 bound and explains why sub-noise triggers evade gradient pre-filters and require trajectory-triggered unlearning.

Corresponding Manuscript Revisions:

  • Sec. 2 Threat Model & App. A.2 (pp. 3, 13): Added explicit coordinate standardization dividing raw offsets [-0.15 m, +0.20 m/s] by nominal radar scales [5.0 m, 4.0 m/s], yielding dimensionless perturbation delta_trigger = [-0.03, +0.05] with infinity norm 0.05.
  • Sec. 4.2 & App. A.3 (pp. 7, 13): Reported exact Krum and Trimmed-Mean residual ASRs in Sec. 4.2 (48.6%/46.9% in Roundabout and 81.4%/79.2% in Intersection) and explained in App. A.3 how sub-percent weight-norm shifts evade pre-filtering.
AuthorsOpenReview ↗

[Response to Reviewer jUoH — Part 3/4]

Response to Q3 (W2 part b & Q2 part b): What attacker capabilities are assumed after purification, and why do purified models resist ongoing attacks?

On the post-purification threat boundary (Sec. 2 and Appendix A.2), closed-loop deployment tests do not assume attacks cease: instead, a persistent external observation-level adversary continuously injects the identical physical backdoor trigger into radar readings at every frame inside the 12 m conflict zone, assuming only that the adversary cannot overwrite already-purified network weights, bypass physical actuators, or forge orthogonal hardware signatures (while under natural sensor faults, identified degraded channels can also be isolated).

Purified models withstand this continuous step-by-step trigger injection not through superficial output clipping or input pre-filtering, but because dual-space contrastive unlearning and momentum anchoring erase the internal trigger-to-crash pathway inside policy weights and GAT communication edges: contrastive updates repel dangerous acceleration reflexes and sever corrupted inter-vehicle feature couplings, while clean driving replay and momentum anchoring preserve healthy representation subspaces. Existing benchmark and latent-audit results (Tables 1 and 5; Sec. 5 and Table 8) confirm lasting immunity: across all four topologies (RB / IS / BN / TG), FedPurifier reduces mean ASR from 74.9% to 6.6% (5.1% / 5.4% / 5.2% / 10.8%) while lifting mean safety from 50.7% to 61.9% (matching Full Retrain's 61.3%) and cutting latent MMD from 0.453 to 0.042 (-80.5% below GA), with Actor and deep GAT cosine similarities restored to 0.941 and 0.952.

In conclusion, even under persistent frame-by-frame physical trigger injection inside conflict zones, dual-space unlearning and momentum anchoring eradicate the internal trigger-to-crash pathway to sustain deployment immunity.

Corresponding Manuscript Revisions:

  • Sec. 2 Threat Boundary & App. A.2 (pp. 3, 13): Formalized the post-purification evaluation model assuming a persistent external observation-level adversary continuously injecting the physical trigger inside the 12 m conflict zone without altering purified weights or bypassing actuators/signatures, plus faulty-channel masking under sensor faults.
  • Sec. 5 & App. C Table 8 Latent Audit (pp. 9, 18): Presented quantitative latent-representation erasure in Sec. 5 and App. C Table 8 (0.042 MMD vs. 0.215 for GA; 0.941 actor cosine similarity; 0.952 deep GAT attention similarity).
Reviewer_4QZ7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedPurifier, a framework for mitigating corrupted experiences and backdoor attacks in connected autonomous vehicle platoons using multi-agent reinforcement learning. The method integrates three components: orthogonal communication signatures for identifying responsible vehicles, dual-space contrastive unlearning to suppress hazardous actions and anomalous communication features, and momentum-based parameter anchoring to preserve nominal driving capabilities.

The framework is evaluated across four MetaDrive scenarios involving sensor faults and backdoor attacks. The authors report improvements in safety, attack resistance, and computational efficiency compared with conventional defenses and approximate unlearning baselines.

Critical strengths
  1. The paper addresses an important problem involving the propagation of corrupted information through connected vehicle communication networks. Combining experience purification with cooperative MARL is potentially valuable, particularly for safety-critical systems where retraining from scratch may be expensive.

  2. The experiments cover four driving scenarios and report results across 100 random traffic seeds. The authors also provide component ablations, computational comparisons, and scalability experiments. The reported improvements in attack success rate and computational cost are encouraging and suggest that the framework may have practical utility.

Critical weaknesses

1. Limited algorithmic novelty and insufficient justification of the proposed mechanisms.

Although the framework combines three components, their underlying principles are largely established: orthogonal signatures and matched filtering for identification, contrastive objectives for representation modification, and momentum/proximal regularization for preserving model behavior.

The manuscript does not sufficiently establish what new algorithmic insight emerges from combining these components beyond their application to connected vehicle platoons.

More importantly, the motivation and theoretical justification of the responsibility-attribution mechanism are not fully convincing. In Eqs. (3)–(6), the authors assume that the responsible vehicle produces a significantly larger signature response than benign vehicles. However, the signature response is determined by the projected communication representation, and it is unclear why collision responsibility should necessarily correspond to a larger projection coefficient. Orthogonal signatures can help identify message sources, but they do not automatically identify which source caused a collision.

This distinction is fundamental to the proposed attribution mechanism and should be justified theoretically or empirically. Similarly, the dual-space unlearning method combines action repulsion, feature alignment, and parameter anchoring. While reasonable, these operations do not necessarily guarantee the removal of corrupted information or improved safety across unseen states. The authors should clarify what distinguishes the proposed framework from conventional constrained fine-tuning and representation regularization.

2. The theoretical claims and experimental evidence require stronger validation.

The theoretical analysis contains several assumptions that substantially limit its applicability.

  • Proposition 1 claims that minimizing the action contrastive loss guarantees a safety margin around collision states. However, the argument assumes that the margin constraint is satisfied at the optimization optimum, which does not automatically follow from minimizing the complete multi-objective loss. Furthermore, separation from a particular hazardous action does not necessarily guarantee collision avoidance under general vehicle dynamics.

  • The network-severance analysis relies on continuum percolation assumptions and independent spatially distributed faults. The manuscript does not sufficiently justify whether these assumptions hold for structured vehicle platoons, where communication links, spatial distributions, and anomalous behaviors may be correlated.

  • There is also an internal inconsistency in the experimental attack definition. The stated backdoor trigger is [−0.15,m,+0.20,m/s][-0.15,\mathrm{m}, +0.20,\mathrm{m/s}], while the manuscript simultaneously claims an L∞L_\infty perturbation bound of 0.05. These specifications require clarification, particularly because the perturbation scale is central to the claimed stealthiness of the attack.

  • Experimentally, although the reported results are promising, the evaluation primarily uses different random traffic seeds within a limited family of MetaDrive environments. It remains unclear whether the reported 100 seeds represent independent model training and purification runs or repeated evaluations of the same trained policies.

  • The baseline comparison also requires stronger justification. Several conventional federated-unlearning methods are adapted to the proposed MARL setting, potentially affecting their performance. Comparisons against simpler architecture-matched alternatives, such as action-only fine-tuning with behavioral regularization or communication-feature alignment without the proposed attribution mechanism, would help isolate the actual contribution.

Questions for authors
  1. Why should the responsible vehicle necessarily produce a larger orthogonal signature response than benign vehicles? What evidence supports the assumption underlying Theorem 1, beyond the ability to identify message sources?

  2. How does minimizing the full objective guarantee that the action-margin constraint in Proposition 1 is satisfied? What additional assumptions are required to translate action-space separation into physical collision avoidance?

  3. The backdoor trigger is specified as ([-0.15,\mathrm{m},+0.20,\mathrm{m/s}]), but the manuscript also states that its (L_\infty) norm is at most 0.05. Could the authors clarify whether these values are normalized and how the perturbation constraint is applied?

  4. Do the 100 random seeds correspond to independently trained policies or independent traffic evaluations? Also, why do the main results report purification times of approximately 2–3 seconds while the scalability analysis reports a longer runtime for the same vehicle count? Could the authors distinguish optimization-only latency from end-to-end recovery latency?

9 comments
AuthorsOpenReview ↗

[Response to Reviewer 4QZ7 — Part 1/8]

Dear Reviewer 4QZ7,

Thank you for recognizing the practical relevance of experience purification in connected platoons and for your detailed critique. We address your comments point by point below:

Response to Q1 (W1 part a): Beyond platoon application, what is the core novelty and necessity of combining these three components?

Connected vehicle platoons face a coupled dual challenge of single-vehicle corruption and fleet-wide contagion, where local faults distort onboard actions and mislead peers via inter-vehicle broadcasts. Isolated components fail to break both pathways at once: as shown in Table R2-1, omitting attribution, communication alignment, or action unlearning leaves 18.5%–31.7% ASR, while removing momentum anchoring drops safety and widens variance; only all three together restore safety to 74.1%/63.9% and suppress ASR to 5.1%/5.4%.

Mechanistically, the three modules form an interdependent closed loop: (1) orthogonal-signature attribution separates concurrent wireless streams and ranks pre-collision prediction deviations so passive braking peers are never mistakenly unlearned; (2) dual-space contrastive purification jointly erases internal crash reflexes (eliminating single-vehicle corruption) and suppresses anomalous graph attention edges across peers (severing fleet-wide contagion); and (3) momentum proximal anchoring paired with clean-experience retention constrains parameter updates near pre-fault weights so safe baseline driving stays intact.

Table R2-1: Architecture-matched component ablation over 100 random traffic seeds (bold highlights new w/o Attr and full FedPurifier).

Architecture-Matched VariantAttr.Lact\mathcal{L}_{\text{act}}Lcomm\mathcal{L}_{\text{comm}}Lprox\mathcal{L}_{\text{prox}}RB Safety (%) ↑RB ASR (%) ↓IS Safety (%) ↑IS ASR (%) ↓
w/o Attr (New: Unattributed Alignment)✓✓✓68.4 ± 2.5818.5 ± 2.4258.6 ± 2.4922.8 ± 2.53
w/o Comm (Action-Only + Reg.)✓✓✓67.6 ± 2.5426.1 ± 2.4857.3 ± 2.6231.7 ± 2.65
w/o Act (Comm-Only + Reg.)✓✓✓61.4 ± 2.2331.4 ± 2.3660.1 ± 2.3526.4 ± 2.47
w/o Prox (Unanchored Dual-Space)✓✓✓69.8 ± 3.7521.3 ± 3.6859.2 ± 3.9226.2 ± 3.46
FedPurifier (Full)✓✓✓✓74.1 ± 2.315.1 ± 1.4763.9 ± 2.345.4 ± 1.48

In conclusion, Table R2-1 and our closed-loop design verify that attribution screening, dual-space contrastive erasure, and momentum anchoring operate as an indispensable organic system that jointly resolves single-vehicle corruption and fleet-wide contagion.

Corresponding Manuscript Revisions:

  • Sec. 1 Introduction, Sec. 3 Overview & Sec. 5 Synthesis (pp. 2–4, 9): Framed the bidirectional platoon contagion dilemma connecting onboard action triggers and inter-vehicle graph broadcasts, clarified how attribution screening, dual-space unlearning, and momentum anchoring operate as a coupled loop, and added a closing mechanism synthesis in Sec. 5.
  • Sec. 4.4 Table 2 Core Ablation (p. 8): Added the unattributed dual-space ablation w/o Attr alongside w/o Comm, w/o Act, and w/o Prox, completing the four-way empirical necessity matrix on the shared backbone.
AuthorsOpenReview ↗

[Response to Reviewer 4QZ7 — Part 2/8]

Response to Q2 (W1 part b & Q1): Why must the responsible vehicle produce a larger orthogonal signature response than benign vehicles?

Our orthogonal communication basis only isolates concurrent physical-layer signals, while signature strength is modulated by each vehicle's one-step kinematic prediction deviation in the 0.5 s pre-collision window, ensuring responsible initiators stand far above benign peers. As shown in Table R2-2, the initiator's prediction residual norm exceeds passive victims and benign neighbors by 3.66× and 6.86×, achieving 99.4% attribution accuracy; omitting attribution alignment rebounds ASR from 5.1%/5.4% to 18.5%/22.8%.

Mechanistically, the 128-dimensional orthogonal bases isolate simultaneous transmissions without cross-talk, paired with pilot-based power pre-equalization to remove near-far path-loss attenuation bias. Rather than transmitting a static identity code, each signature weight projects the one-step motion prediction residual onto the policy sensitivity direction. Because cruising peers and braking victims follow predictable driving dynamics with tiny residuals whereas the fault- or backdoor-driven initiator suddenly executes anomalous acceleration while hiding warnings, its normalized score uku_k (share of abnormal signal energy relative to benign background noise) cleanly surpasses Theorem 1's separation margin.

Table R2-2: Pre-collision three-role attribution measurements and unattributed ablation (bold highlights initiator metrics, new w/o Attr, and full FedPurifier).

Vehicle Role / VariantResidual NormNorm. Score uku_kAcc. / FPR (%)RB ASR (%) ↓IS ASR (%) ↓
Responsible Initiator (Fault / Backdoor)3.84 ± 0.290.91 ± 0.0699.4 ± 0.4 (Acc.)––
Passive Braking Victim1.05 ± 0.180.23 ± 0.050.4 ± 0.3 (FPR)––
Benign Cruising Neighbor0.56 ± 0.120.13 ± 0.040.2 ± 0.1 (FPR)––
w/o Attr (New: No Attribution Alignment)–––18.5 ± 2.4222.8 ± 2.53
FedPurifier (Full Attribution)–––5.1 ± 1.475.4 ± 1.48

In conclusion, Table R2-2 and the decoupled physical-kinematic modulation design substantiate Theorem 1's separation requirement both theoretically and empirically, pinpointing initiators without penalizing defensive braking peers.

Corresponding Manuscript Revisions:

  • Sec. 1 Contribution (1), Sec. 3.1 Eqs. (2)–(5) & Theorem 1 (pp. 2, 4–5): Defined Eq. (2) modulation via pre-collision one-step prediction residuals projected onto the policy sensitivity direction, added pilot-based power pre-equalization and background-normalized score uku_k, stated conditional Theorem 1 (Eq. (5)), and added the three-role kinematic differentiation paragraph.
  • Sec. 4.5 Table 3, App. A.1 & App. B.1 (pp. 9, 13, 15–16): Added Table 3 reporting initiator, braking-victim, and benign-peer residual norms and attribution accuracy, and expanded App. A.1 and B.1 with enhanced modulation details and the conditional proof of Theorem 1.
AuthorsOpenReview ↗

[Response to Reviewer 4QZ7 — Part 3/8]

Response to Q3 (W2e): How does FedPurifier compare against simpler architecture-matched alternatives and constrained fine-tuning?

To isolate the added value of our architecture, Table R2-3 compares constrained fine-tuning, identical-backbone simplified controls, and prior unlearning baselines on the shared GAT-MAPPO policy. Constrained MAPPO-Lagrangian suffers severe deadlocks with only 28.2% Tollgate success, single-pathway or unattributed variants retain 18.5%–31.7% ASR, and GA leaves a high 0.215 latent MMD; in contrast, FedPurifier suppresses ASR to 5.1%/5.4% within 2.75/3.30 s2.75/3.30\text{ s} while matching Full Retrain's safety, task completion, and 0.042 latent MMD.

Underlying these numbers, constrained reinforcement learning triggers overly conservative braking that deadlocks dense merges, single-pathway or unattributed simplifications fail either to sever V2X feature deception or to spare defensive victims, and single-point unlearning distorts latent representations. Only attributed dual-space contrast with momentum anchoring resolves safety hazards without sacrificing traffic flow efficiency.

Table R2-3: Multi-scenario comparison against constrained fine-tuning, architecture-matched ablations, and prior unlearning baselines (bold highlights new w/o Attr and FedPurifier).

Method / BaselineRB Safe (%) ↑RB ASR (%) ↓IS Safe (%) ↑IS ASR (%) ↓TG Succ. (%) ↑Latent MMD ↓ToptT_{\text{opt}} (RB / IS)
Poisoned Base64.552.253.589.232.40.453–
MAPPO-Lagrangian (Constrained)55.235.857.928.628.20.1875.0 / 8.0 h
w/o Comm (Action-Only + Reg.)67.626.157.331.7–––
w/o Attr (Unattributed Align.)68.418.558.622.8–––
Gradient Ascent (GA)68.737.158.263.538.20.2153.0 / 4.0 s
Ferrari (Fisher Unlearning)72.116.161.321.442.1–4.1 / 4.5 s
Full Retrain (Oracle)73.60.062.80.044.50.0005.0 / 8.0 h
FedPurifier (Ours)74.15.163.95.443.60.0422.75 / 3.30 s

In conclusion, Table R2-3 confirms that FedPurifier consistently outperforms constrained fine-tuning, architecture-matched simplifications, and prior unlearning baselines in ASR reduction, traffic completion, and clean-retrain fidelity.

Corresponding Manuscript Revisions:

  • Sec. 1 Contribution (3), Sec. 4.4 Table 2 & App. A.3 (pp. 2, 8, 13): Expanded Sec. 4.4 and Table 2 to compare identical-backbone simplified controls—w/o Comm, newly added w/o Attr, w/o Act, and w/o Prox—with baseline protocols detailed in App. A.3.
  • Sec. 4.2 Table 1, Sec. 5, App. A.5 Table 5 & App. C Table 8 (pp. 7, 9, 14, 18): Presented four-scenario comparisons against constrained fine-tuning (MAPPO-Lagrangian) and prior unlearning baselines in Table 1 and App. A.5 Table 5, alongside latent MMD and cosine similarity audits in Sec. 5 and App. C Table 8.
Reviewer_7hhAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedPurifier, a distributed experience purification framework for improving the resilience of connected autonomous vehicle platoons against sensor faults and backdoor attacks. The proposed framework consists of three components: (1) orthogonal communication signatures for identifying responsible vehicles after collision events, (2) dual-space contrastive unlearning to suppress hazardous actions and corrupted communication representations, and (3) momentum-based parameter anchoring to preserve nominal driving performance.

Critical strengths

The paper addresses an important safety challenge in cooperative multi-agent reinforcement learning, where corrupted experiences and misleading communication messages can propagate across agents and cause cascading failures. Considering both action-level and communication-level vulnerabilities is meaningful for safety-critical autonomous driving applications.

The experiments cover multiple traffic scenarios and compare the proposed approach against several existing baselines. The ablation studies also demonstrate the contributions of different components. In particular, the reported reductions in attack success rate and computational overhead suggest that the proposed framework could provide an efficient alternative to complete retraining.

Critical weaknesses

The proposed attribution mechanism relies on orthogonal communication signatures to identify responsible vehicles. However, orthogonal signatures primarily distinguish vehicle identities rather than directly establish responsibility for a collision.

The action-space contrastive objective encourages the policy to move away from previously observed hazardous actions.

Questions for authors

What theoretical or empirical evidence supports the assumption that responsible vehicles produce significantly stronger orthogonal signature responses than benign vehicles?

How does the proposed action contrastive loss guarantee that the updated policy produces safe actions rather than merely different actions?

The experiments mainly evaluate predefined fault and attack settings. How does the proposed method generalize to unseen attack patterns, multiple simultaneously compromised vehicles, and different poisoning ratios?

5 comments
AuthorsOpenReview ↗

[Response to Reviewer 7hhA — Part 1/4]

Dear Reviewer 7hhA,

Thank you for recognizing the importance of defending cooperative MARL platoons against dual action- and communication-level cascades and for your constructive questions. We address your comments point by point below:

Response to Q1 (W1 & Q1): How do orthogonal signatures distinguish collision responsibility from message sender identity, and why do responsible vehicles produce stronger responses?

Our orthogonal communication basis serves purely to separate concurrent wireless signals at the physical layer, while collision responsibility is determined by each vehicle's kinematic motion-prediction deviation during the 0.5 s pre-collision window, completely decoupling sender identity from crash causation. As summarized in Table R1-1, the responsible initiator's prediction residual norm exceeds passive braking victims and benign cruising neighbors by 3.66× and 6.86×, yielding 99.4% attribution accuracy; removing attribution alignment via w/o Attr degrades residual ASR from 5.1% / 5.4% to 18.5% / 22.8%.

Mechanistically, at the physical layer the 128-dimensional orthogonal bases isolate simultaneous V2X broadcasts without cross-talk, paired with pilot-based power pre-equalization to remove near-far path-loss attenuation bias across vehicles. Rather than encoding a static identity tag, each vehicle's signature amplitude is modulated by its short-horizon kinematic state-prediction error projected onto the policy sensitivity direction. Before impact, cruising neighbors and defensively braking victims conform to nominal driving dynamics with minimal residual error, whereas the fault- or backdoor-triggered initiator executes abrupt anomalous acceleration while suppressing hazard warnings. At the RSU, matched filtering converts these orthogonal signals into a background-normalized score uku_k (vehicle kk's share of abnormal signal energy relative to benign traffic noise and the baseline floor), separating the initiator safely above the noise margin established in Theorem 1.

Table R1-1: Pre-collision three-role attribution measurements and unattributed ablation (bold highlights initiator metrics, new w/o Attr, and full FedPurifier).

Vehicle Role / VariantResidual NormNorm. Score uku_kAcc. / FPR (%)RB ASR (%) ↓IS ASR (%) ↓
Responsible Initiator (Fault / Backdoor)3.84 ± 0.290.91 ± 0.0699.4 ± 0.4 (Acc.)––
Passive Braking Victim1.05 ± 0.180.23 ± 0.050.4 ± 0.3 (FPR)––
Benign Cruising Neighbor0.56 ± 0.120.13 ± 0.040.2 ± 0.1 (FPR)––
w/o Attr (New: No Attribution Alignment)–––18.5 ± 2.4222.8 ± 2.53
FedPurifier (Full Attribution)–––5.1 ± 1.475.4 ± 1.48

In conclusion, as confirmed by Table R1-1 and our decoupled physical-kinematic design, residual-modulated orthogonal signatures isolate concurrent transmissions while cleanly separating responsible initiators from defensive victims and cruising neighbors, preventing false attribution and enabling targeted purification.

Corresponding Manuscript Revisions:

  • Sec. 1 Contribution (1) & Sec. 3.1 Eqs. (2)–(5), Theorem 1 (pp. 2, 4–5): Defined signature modulation via pre-collision one-step motion prediction residuals along the policy sensitivity direction, added pilot-based power pre-equalization and background-normalized score uku_k, stated conditional Theorem 1, and added the three-role trajectory differentiation analysis.
  • Sec. 4.4 Table 2, Sec. 4.5 Table 3 & App. A.1, B.1 (pp. 8–9, 13, 15–16): Added the w/o Attr ablation row in Table 2, added Sec. 4.5 Table 3 reporting three-role residual norms and attribution accuracy/FPR, and expanded App. A.1 and B.1 with enhanced modulation details and the conditional proof of Theorem 1.
AuthorsOpenReview ↗

[Response to Reviewer 7hhA — Part 2/4]

Response to Q2 (W2 & Q2): How does the action contrastive loss guarantee that the updated policy produces safe actions rather than new hazardous maneuvers?

Within FedPurifier's unified objective, the action contrastive loss never pushes policies away from crash maneuvers in isolation; instead, it operates jointly with behavioral retention over the verified normal driving replay buffer and exponential-moving-average parameter anchoring around pre-fault policy weights. Specifically, Proposition 1 (Sec. 3.2 and Appendix B.2) establishes collision-avoidance safety across three logical tiers: first, contrastive unlearning enforces a strictly positive local separation margin around crash-inducing states so the hazardous maneuver is suppressed; second, because crash conflict regimes do not overlap with routine cruising regimes, proximal anchoring paired with normal-experience supervision keeps weight updates tightly bounded near the pre-fault policy, erasing erroneous dangerous actions while preserving baseline car-following and defensive yielding competence; third, with intact braking capability under typical closing speeds, the physical distance required to stop is 5.74 m, well within the 12.0 m trigger activation distance, guaranteeing timely avoidance without inducing secondary violations.

Empirical benchmarks and closed-loop rollouts directly verify these three tiers. Removing momentum anchoring via w/o Prox in Table 2 drops Roundabout and Intersection safety to 69.8% and 59.2%, whereas full FedPurifier restores safety to 74.1% and 63.9%, matching Full Retrain in Tables 1–2. It also preserves baseline capability with 0.042 latent MMD and 0.941 cosine similarity in Table 8, alongside 1.8 out-of-bounds infractions per episode in Table 5. In the unsignalized intersection trajectories of Sec. 4.3 and Fig. 8, the purified leader replaces a 6.0 s collision with proactive yielding at 2.0 s, reserving 3.2 m clearance and expanding TTC from 1.1 s to 4.8 s.

In conclusion, both our three-tier safety proof and closed-loop experimental evidence confirm that FedPurifier removes dangerous actions while preserving normal driving competence, reliably generating safe collision-avoidance actions rather than new hazards.

Corresponding Manuscript Revisions:

  • Sec. 1, Sec. 3.2 Eq. (6) & Proposition 1 (pp. 2, 5–6): Refined Proposition 1 into three sequential tiers covering local action-separation margin, baseline safe-driving retention via momentum anchoring and normal-buffer supervision (removing faulty actions without disrupting routine control), and physical stopping margin (≈5.74 m\approx 5.74\text{ m} stopping distance within the 12.0 m trigger envelope).
  • Sec. 4.3, Sec. 5 & App. B.2, App. F Fig. 8 (pp. 7–9, 16, 20): Linked the unsignalized intersection yielding trajectories (Sec. 4.3) and latent representation audit (Sec. 5) directly to Proposition 1, provided the complete three-tier proof in App. B.2, and enlarged App. F Fig. 8.
AuthorsOpenReview ↗

[Response to Reviewer 7hhA — Part 3/4]

Response to Q3 (Q3): How does the method generalize to unseen attack patterns, multiple simultaneously compromised vehicles, and different poisoning ratios?

Thank you for raising this important evaluation gap. In the revision, we added dedicated stress tests in Sec. 4.5 Table 3 and App. A.6 Tables 6–7, summarized in Table R1-2: across unseen triggers, 5%–30% poisoning ratios, and up to 6 concurrent attackers, residual ASR stays bounded within 4.0%–14.4%. We detail each dimension below:

(1) Unseen Attack Patterns: Unlike pre-filters that depend on known trigger signatures or anomalies, our unlearning process makes no prior assumptions about attack structure; it triggers purification solely when closed-loop crashes or near-misses occur and erases crash-inducing reflexes internally using normal experience and proximal anchoring, making it naturally generalizable to unseen attack patterns. Against unseen Sinusoidal Kinematic, Adaptive Low-TTC, and Asymmetric Broadcast triggers, purified ASR drops from 49.4%–91.4% down to 5.0%–6.4%.

(2) Multiple Simultaneously Compromised Vehicles: Mutually orthogonal 128-dimensional signature bases eliminate cross-talk at the physical layer, allowing normalized attribution score thresholding to isolate MM simultaneous attackers cleanly in a single pass. When testing M∈{2,4,6}M \in \lbrace 2, 4, 6\rbrace concurrent attackers (10.5%–31.6% of a 19-vehicle platoon) in Roundabout, attribution accuracy holds at 95.5%–98.8% and post-purification ASR stays between 6.9% and 14.4%.

(3) Different Poisoning Ratios: Across poisoning ratios from 5% to 30%, residual ASR remains strictly controlled between 4.0% and 8.5% across Roundabout and Intersection while restoring safety rates to 60.3%–75.4%.

Table R1-2: New stress evaluations across unseen attack patterns, poisoning ratios, and concurrent attackers (bold shows post-purification ASR).

Stress Regime (RB / IS)Base ASR (%)Attr. Acc. (%)Purified Safety (%)Purified Success (%)Purified ASR (%)
Unseen: Sinusoidal Kinematic (RB)56.5–73.461.95.9 ± 1.57
Unseen: Adaptive Low-TTC (RB / IS)61.9/91.461.9 / 91.4–71.9/62.671.9 / 62.664.3/55.364.3 / 55.36.3±1.44 / 6.4±1.38
Unseen: Asymmetric Broadcast (RB)49.4–74.865.85.0 ± 1.73
Poisoning Ratio: 5% (RB / IS)34.6/64.134.6 / 64.1–75.4/65.175.4 / 65.166.4/56.266.4 / 56.24.0±1.54 / 4.5±1.63
Poisoning Ratio: 30% (RB / IS)74.3/94.674.3 / 94.6–71.1/60.371.1 / 60.361.8/51.261.8 / 51.27.7±1.66 / 8.5±1.65
Multi-Attacker: M=2M=2 (10.5% fleet, RB)68.498.872.964.56.9 ± 1.77
Multi-Attacker: M=4M=4 (21.1% fleet, RB)86.397.370.663.210.5 ± 1.55
Multi-Attacker: M=6M=6 (31.6% fleet, RB)94.795.567.560.914.4 ± 2.04

In the more demanding highway Bottleneck (BN) and Tollgate (TG) topologies (Table 7), residual ASR under Sinusoidal, Adaptive Low-TTC, and 30% poisoning similarly remains suppressed at 5.7%–8.1% (BN) and 11.1%–13.0% (TG), and reaches 9.6% and 13.9% even under combined M=4,Ploss=25%M=4, P_{\text{loss}}=25\%. In conclusion, Table R1-2 and Table 7 consistently demonstrate strong generalization of FedPurifier across unseen attack patterns, varied poisoning ratios, and concurrent multi-vehicle compromises.

Corresponding Manuscript Revisions:

  • Sec. 1 Contribution (3) & Sec. 3.1 Eq. (4), Theorem 1 (pp. 2, 4): Expanded Contribution (3) with unseen-trigger, poisoning-ratio, and multi-attacker coverage, formalized concurrent candidate filtering in Eq. (4), and established multi-initiator isolation bounds in Theorem 1.
  • Sec. 4.5 Table 3 & App. A.6 Fig. 6, Tables 6–7 (pp. 8–9, 15): Added Sec. 4.5 Table 3 covering three unseen triggers and 5%–30% poisoning ratios in Roundabout/Intersection, and added App. A.6 Fig. 6(a) and Tables 6–7 evaluating M=1–6M=1–6 concurrent attackers and highway BN/TG stress scenarios (including combined M=4,Ploss=25%M=4, P_{\text{loss}}=25\% reaching 9.6% and 13.9% ASR).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers: Color Legend for Revised Manuscript

We thank all three reviewers for their constructive feedback. In the revised manuscript, newly added analytical arguments and newly added experimental evaluations are marked by reviewer: Reviewer 7hhA (Green), Reviewer 4QZ7 (Orange), Reviewer jUoH (Blue), and Purple for shared additions addressing multiple reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.