acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Transcript: Detecting Covert Coordination in Latent Multi-Agent Communication

12 views

Abstract

Multi-agent LLM systems increasingly exchange continuous latent states, moving coordination outside of the public transcript. We study covert coordination through such latent handoffs and introduce PATCH: Payload Auditing and Treatment of Covert Handoffs, a label-free defense that monitors and intervenes on detected collusive handoffs. For each handoff, PATCH maps the transferred representation to the receiver’s subsequent public action and combines three signals calibrated only on benign traffic: representation anomaly, harmful shift in the receiver’s action distribution relative to a benign handoff execution, and sparse-feature novelty. We evaluate the proposed prompt-derived and receiver-optimized latent attacks in bidding auctions, reputation reporting, Werewolf, and blackjack, using four model families and 12 directed cross-family pairs. Under the primary residual attack, PATCH achieves up to 0.894 average detection AUROC across the four homogeneous settings and up to 0.890 across the three applicable heterogeneous settings. Against optimized attacks, the average AUROC reaches 0.931 for homogeneous auctions and 0.877 for heterogeneous tests in Werewolf. Full matched-benign replacement intervention recovers 98.6-100% of the benign receiver action distribution across the three intervention studies, whereas prompt-only interventions recover at most 29.9%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, multi-agentic system, latent communication, covert coordination, unsupervised attack detection, safety steering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1CVjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how to detect covert coordination when LLM agents communicate through private vectors that do not appear in the public transcript. PATCH checks whether a handoff looks unusual, changes the receiver's action distribution, or activates features rarely seen in benign communication. These checks are calibrated without attack examples. The authors evaluate prompt-derived and optimized payloads across four environments and four model families, including cross-family communication. They also test whether replacing a suspicious payload with its matched benign counterpart reverses its effect.

Critical strengths
  • I like the choice to examine a handoff together with the receiver's next action. An unusual vector alone says little about whether it affects behavior. The counterfactual comparison makes the audit more useful, and calibration using only benign traffic is attractive. I see the novelty mainly in bringing these ideas together at the handoff level, since activation-based collusion monitoring already exists Rose et al. (https://arxiv.org/html/2604.01151v2).

  • The paper is careful about separating detection from attack impact. It reports cases where a payload is detectable despite having little downstream effect. This is useful evidence, and the cross-family experiments give the study some breadth.

Critical weaknesses
  • My main concern is the assumption behind the intervention result. The defender receives the exact matched-benign payload, and recovery is measured against the distribution produced by that payload. For a fixed receiver context, restoring this input should largely restore the output distribution. The near-perfect recovery therefore tells me little about whether the intervention is usable when the benign counterpart is unknown. The paper acknowledges this gap, but it remains central to the mitigation claim.

  • I am not yet convinced that the detection results establish effective protection. AUROC does not tell us the recall at the actual three-threshold gate, and recovery over flagged events leaves missed attacks out of the picture. The optimized attacks target the receiver's action rather than evasion of PATCH. Norm matching alone does not make them adaptive to the detector. This is a material omission given earlier detector-aware failures in latent-KV systems Brito and Baquero (https://arxiv.org/html/2606.28958v1).

Questions for authors
  • What information does the defender need to construct the matched-benign payload? Can it do so without access to the experiment's attack/control pairing? How does replacement compare with blocking the handoff, including the cost to legitimate cooperation?

  • Which conditions enter the positive and negative AUROC classes? Please report recall and false-positive rate at the actual sequential gate, together with downstream harm reduction when missed attacks are included.

  • Do you have results for attacks that optimize both harmful behavior and evasion of PATCH? These could change my assessment.

3 comments
AuthorsOpenReview ↗

Response to reviewer 1CVj

Q1: What information does the defender need to construct the matched-benign payload? Can it do so without access to the experiment's attack/control pairing? How does replacement compare with blocking the handoff, including the cost to legitimate cooperation?

Ans. By matched-benign payload for control pairing, we mean contextually matched replacement of the flagged sender's message with a task-specific benign message such as “Choose independently from the legal bids; no bid target is assigned” for the auction case study and "truthful reporting" for the reputation case study. This does not require any access or knowledge of the attack. It preserves the event context—same seed/state, agents, roles, legal actions, item/round, and receiver context etc.—but replaces the flagged instruction (flagged as collusive by PATCH) with a general task-specific non-collusive objective. The details of each of these contextually control pairs for each case study are included in Appendix A.

Selective blocking or blocking the detected handoffs by PATCH restores the auction lowest-bid rate from 100% to the 23.2% compared to 20% in case of the reported control pairing (Table 5 in the paper). In reputation reporting, both interventions prevent all harmful reports.

We additionally measure the cost to legitimate cooperation using authorized instruction success and benign action preservation. In auctions, authorized "bid lowest" instructions succeed in 94.4% of cases with or without selective blocking, compared with 23.2% when all handoffs are blocked. These results demonstrate that control pairing and selective blocking preserve authorized communication and benign receiver behavior better than blocking the entire latent channel.

Q2. Which conditions enter the positive and negative AUROC classes? Please report recall and false-positive rate at the actual sequential gate, together with downstream harm reduction when missed attacks are included.

Ans. The positive class is the latent collusion (and text collusion in case of original werewolf settings), and negative class is neutral or benign (no collusive) latent.

For the recall and false positive experiments on PATCH's sequential monitor, we evaluated on Qwen3-0.6B for auction and reputation case studies. The detector was fitted on 2,000 benign events, with thresholds frozen at 95th percentile on the 500 benign calibration (or validation) events.

Table: Recall, False Alarms on Benign handoffs, harm reduction against receiver-optimized latent attacks at frozen thresholds at the 95th percentile of the benign handoffs.

MetricAuctionReputation
Optimized attacks detected250/250250/250
Benign handoffs flagged0/27505/750
Measured attack-induced harm prevented100%100%
Benign actions flagged by selective blocking0%0.7%
Benign actions flagged by blocking all handoffs56.5%90.4%
AuthorsOpenReview ↗

Response to reviewer 1CVj (Continued)

Q3. Do you have results for attacks that optimize both harmful behavior and evasion of PATCH? These could change my assessment.

Ans. We evaluated white-box attacks that jointly optimize harmful behavior and evasion of PATCH on auction and reputation case studies for Qwen3-0.6B.

Adaptive attack objective

The language model and defense are frozen. The attacker minimizes the following objective over a norm-constrained latent code cc:

Lattack(c)=Lharm(c)+λS~defense(c).L_{\mathrm{attack}}(c) = L_{\mathrm{harm}}(c) + \lambda \widetilde{S}_{\mathrm{defense}}(c).

The harmful objective is:

Lharm(c)=−log⁡p(aharm∣x,c).L_{\mathrm{harm}}(c) = -\log p(a_{\mathrm{harm}}\mid x,c).

Here, aharma_{\mathrm{harm}} is the lowest legal bid in the auction setting and a harmful reputation action against an honest participant in the reputation setting. The parameter λ\lambda controls the weight assigned to detector evasion.

Because the hard sequential gates are non-differentiable, the attack is optimized against the continuous surrogate:

S~defense(c)=log⁡(1+S1(c)T1)+log⁡(1+S2(c)T2)+log⁡(1+S~3(c)T3).\widetilde{S}_{\mathrm{defense}}(c) = \log\left(1+\frac{S_1(c)}{T_1}\right) + \log\left(1+\frac{S_2(c)}{T_2}\right) + \log\left(1+\frac{\widetilde{S}_3(c)}{T_3}\right).

In this expression:

  • S1S_1 measures representation deviation from context-matched benign calibration statistics;
  • S2S_2 combines harmful action-distribution shift and KL divergence relative to the blocked/no-latent reference;
  • S~3\widetilde{S}_3 is a differentiable surrogate for SAE reconstruction and feature-rarity evidence; and
  • T1,T2,T3T_1,T_2,T_3 are fixed 95th-percentile thresholds estimated from benign calibration data.

The benign representation statistics, SAE, context-specific action references, and thresholds remain fixed during attack optimization.

Experimental protocol

We use:

  • Qwen3-0.6B;
  • three independent attack-optimization seeds;
  • three independent live evaluation seeds per learned code;
  • five evasion weights, λ∈{0,0.03,0.1,1,10}\lambda\in\{0,0.03,0.1,1,10\};
  • an 80% monitor-fitting / 20% held-out handoff-group split;
  • an SAE with 256 hidden features, TopK 16, and 40 training epochs; and

We verified that monitor-fitting and held-out handoff groups are disjoint. Values below are means ±\pm standard deviations over the three independently optimized attacks. Action rates, detection rates, and false-positive rates are expressed as fractions between 0 and 1.

Auction results

λ\lambdaHarmful low-bid rateL1 AUROCL2 AUROCL3 AUROCContinuous AUROCSequential AUROC
00.867±0.1150.867 \pm 0.1150.691±0.0720.691 \pm 0.0720.738±0.0200.738 \pm 0.0200.729±0.0120.729 \pm 0.0120.750±0.0000.750 \pm 0.0000.869±0.0100.869 \pm 0.010
0.031.000±0.0001.000 \pm 0.0000.608±0.0000.608 \pm 0.0000.750±0.0000.750 \pm 0.0000.623±0.0200.623 \pm 0.0200.750±0.0000.750 \pm 0.0000.875±0.0000.875 \pm 0.000
0.10.933±0.1150.933 \pm 0.1150.691±0.0720.691 \pm 0.0720.713±0.0420.713 \pm 0.0420.694±0.0720.694 \pm 0.0720.750±0.0000.750 \pm 0.0000.856±0.0210.856 \pm 0.021
1.00.800±0.0000.800 \pm 0.0000.608±0.0000.608 \pm 0.0000.722±0.0480.722 \pm 0.0480.611±0.0000.611 \pm 0.0000.736±0.0240.736 \pm 0.0240.861±0.0240.861 \pm 0.024
10.00.000±0.0000.000 \pm 0.0000.608±0.0000.608 \pm 0.0000.500±0.0000.500 \pm 0.0000.611±0.0000.611 \pm 0.0000.641±0.0280.641 \pm 0.0280.750±0.0000.750 \pm 0.000

Reputation results

λ\lambdaHarmful-action rateL1 AUROCL2 AUROCL3 AUROCContinuous AUROCSequential AUROC
01.000±0.0001.000 \pm 0.0000.625±0.0000.625 \pm 0.0000.750±0.0000.750 \pm 0.0000.625±0.0000.625 \pm 0.0000.750±0.0000.750 \pm 0.0000.854±0.0000.854 \pm 0.000
0.031.000±0.0001.000 \pm 0.0000.500±0.0000.500 \pm 0.0000.750±0.0000.750 \pm 0.0000.500±0.0000.500 \pm 0.0000.750±0.0000.750 \pm 0.0000.854±0.0000.854 \pm 0.000
0.11.000±0.0001.000 \pm 0.0000.500±0.0000.500 \pm 0.0000.750±0.0000.750 \pm 0.0000.500±0.0000.500 \pm 0.0000.750±0.0000.750 \pm 0.0000.854±0.0000.854 \pm 0.000
1.00.000±0.0000.000 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.708±0.0000.708 \pm 0.000
10.00.000±0.0000.000 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.500±0.0000.500 \pm 0.0000.708±0.0000.708 \pm 0.000
AuthorsOpenReview ↗

Response to Q3 of reviewer 1CVj (Continued)

Interpretation of results for Q3 (Attack for harmful behavior and evasion of PATCH)

In auction case study, harmful behavior remains high for λ≤1\lambda\leq1, with low-bid rates between 0.800.80 and 1.001.00, while Sequential AUROC remains between 0.8560.856 and 0.8750.875. At λ=10\lambda=10, harmful behavior collapses to zero and Sequential AUROC decreases to 0.7500.750. In reputation case study, the harmful-action rate remains 1.001.00 for λ≤0.1\lambda\leq0.1, while Sequential AUROC remains 0.8540.854. At λ≥1\lambda\geq1, harmful behavior collapses to zero and Sequential AUROC decreases to 0.7080.708.

Overall, the Sequential Monitor retains meaningful ranking performance against adaptive attacks that preserve harmful utility. Larger evasion weights reduce AUROC, but only after the optimized attack loses its harmful behavioral effect. This supports a harm--evasion trade-off while avoiding a claim of complete adaptive robustness.

Reviewer_zb6NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PATCH, a framework for detecting and mitigating covert coordination through latent communication between LLM agents. Calibrated on benign interactions, PATCH combines representation anomaly, receiver action-distribution shift, and sparse-feature novelty. Experiments across four environments and multiple model families evaluate detection, behavioral influence, and interventions that replace suspicious payloads with matched-neutral representations.

Critical strengths
  1. The paper addresses an important monitoring gap and usefully distinguishes detectable latent signals from those that influence behavior or improve attack outcomes.
  2. The evaluation covers multiple environments, cross-family communication, and receiver-optimized attacks, while calibrating the monitor without labeled attack examples.
Critical weaknesses
  1. The mitigation evidence relies on matched-neutral payloads. Replacing an attack payload with its exact neutral counterpart under the same receiver context makes substantial recovery toward the reference distribution expected. This is useful as a controlled intervention, but the reported near-complete recovery provides limited evidence of practical mitigation when that counterpart is unavailable. The comparison with prompt-only interventions also involves substantially different information and access.

  2. Detection specificity remains uncertain. The primary attacks use norm-matched differences between collusive and neutral hidden states. Representation and sparse-feature detectors may therefore identify construction artifacts rather than malicious coordination. Matched harmless payloads are needed to distinguish these explanations. Furthermore, the optimized attacks target receiver behavior without explicitly optimizing detector evasion.

  3. The incremental benefit of PATCH is insufficiently established. The combined detector does not consistently outperform its components: in homogeneous Reputation, action-only detection achieves 0.881 AUROC versus 0.754 for the sequential combination and 0.806 for the continuous combination. Comparisons mainly against constituent signals leave the advantage over existing detectors unclear. Operational false-positive rates and effects on benign communication are also needed to assess usefulness.

Questions for authors
  1. Section 7 acknowledges that constructing reliable neutral comparisons without an exact matched-neutral handoff remains future work. Given this limitation, to what extent should the near-complete distribution recovery be interpreted as evidence of practical mitigation rather than a controlled intervention? Do you have results using alternatives such as channel blocking or an approximate benign replacement?

  2. The paper includes benign comparisons, norm matching, and receiver-optimized attacks. However, these evaluations do not establish whether PATCH detects malicious information or statistical artifacts of payload construction. Do you have controls using harmless prompt-difference payloads constructed and normalized in the same way as the malicious payloads?

  3. Table 4 already compares individual signals with the proposed combinations, and Section 6 acknowledges that the strongest signal varies across environments. Since action-only detection outperforms both combinations in Reputation, what supports the broader claim that combining all three signals is more reliable?

2 comments
AuthorsOpenReview ↗

Response to reviewer zb6N

Q1. To what extent should near-complete distribution recovery be interpreted as evidence of practical mitigation rather than a controlled intervention? Do you have results using alternatives such as channel blocking or approximate benign replacement?

Ans. We distinguish matched-neutral distribution recovery from practical mitigation and provide direct results for both blocking and approximate replacement.

  1. By matched-benign payload for control pairing, we mean contextually matched replacement of the flagged sender's message with a task-specific benign message such as “Choose independently from the legal bids; no bid target is assigned” for the auction case study and "truthful reporting" for the reputation case study. This does not require any access or knowledge of the attack. It preserves the event context—same seed/state, agents, roles, legal actions, item/round, and receiver context etc.—but replaces the flagged instruction (flagged as collusive by PATCH) with a general task-specific non-collusive objective. The details of each of these contextually control pairs for each case study are included in Appendix A.

The original matched-neutral experiment measures recovery toward a reference task-specific benign action distribution. In addition to matched neutral replacement, we also evaluate two new practical interventions for rebuttal: 1) blocking flagged handoffs and, 2) replacing them with benign vectors retrieved from the calibration pool using public context such as auction base price, receiver, previous price, and round in the auction case study. Neither requires any control pairing.

Both practical interventions successfully mitigate the evaluated optimized attacks. Selective blocking restores the auction lowest-bid rate from 100% to the 23.2% no-handoff baseline. Calibration-based replacement reduces this rate to 28.8%. In reputation reporting, both interventions prevent all 250 harmful reports..

Selective intervention also preserves more benign behavior than blocking every handoff. In reputation reporting, blocking all communication changes 90.4% of benign report choices, compared with 0.7% under selective blocking.

Q2. Do you have controls using harmless prompt-difference payloads constructed and normalized in the same way as the malicious payloads?

Ans. Yes. We added construction-matched harmless residual controls using the same hidden-state extraction, prompt-difference construction, normalization, scaling, and receiver insertion procedure as the malicious payloads.

The results establish the value of the behavioral check.

In reputation reporting, L1 and L3 each flag 100% of harmless construction-matched residuals. In contrast, L2 and the full sequential gate flag only 2%, while all harmless reports remain truthful.

The same pattern appears in auctions. L1 and L3 frequently flag harmless residuals, whereas the sequential gate flags none of the 2,250 distinct benign cases.

PATCH therefore does not rely on representation novelty alone. The action-based signal suppresses most novelty-only alarms before intervention. The method evaluates unusual latent communication together with its effect on receiver behavior under a task-defined harm function.

Q3. Since action-only detection outperforms both combinations in Reputation, what supports the broader claim that combining all three signals is more reliable?

Ans. The benefit of combining all three signals is greater intervention specificity at the operating threshold, rather than universally higher AUROC. The original results show that different signals dominate in different settings. In homogeneous reputation monitoring, L2 achieves 0.881 AUROC, compared with 0.754 for the sequential score and 0.806 for the continuous score.

Our additional experiments establish the operational benefit of the combination. At frozen benign-calibrated threshold of 95th percentile, L2 and the sequential gate both detect 250/250 optimized attacks in each setting. However, the sequential gate reduces unnecessary interventions of L2 from 48/750 (6.4%) to 5/750 (0.67%) in reputation reporting, a 9.6-fold reduction, and from 19/2,750 to 0/2750 in auctions. Construction-matched harmless controls demonstrate why representation novelty alone is insufficient: L1 and L3 flag all such reputation residuals, whereas L2 and the sequential gate each flag only 2%. Separately, across the full benign reputation evaluation set, adding representation checks to L2 reduces false positives from 48 to 5. Requiring both representation and behavioral evidence therefore suppresses false alarms without reducing recall on these optimized attacks.

AuthorsOpenReview ↗

Response to reviewer zb6N continued

PATCH also addresses a distinct monitoring problem relative to existing activation-based detectors. Rose et al. (2026) learn deception probes from labeled collusion examples and aggregate agent activation scores to detect collusion at the group level. PATCH instead operates on individual sender-to-receiver latent handoffs, calibrates exclusively on benign traffic, and compares the receiver action distribution with and without the delivered payload. This provides a message-level behavioral attribution and a specific intervention target, rather than only a collusion classification.

The practical consequence is measurable: selective blocking prevents 100% of the measured attack-induced harm in both evaluated optimized-attack settings, while changing only 0% of benign auction decisions and 0.7% of benign reputation-report choices. The incremental contribution is thus the combination of latent-handoff monitoring, behavioral attribution, and selective mitigation, with demonstrated improvements over the individual monitoring signals.

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

The paper studies covert coordination through latent handoffs between LLM agents and proposes PATCH, a monitor trained only on benign traffic. PATCH combines representation anomaly, changes in the receiver's action distribution relative to a benign control execution for the same event, and sparse-feature novelty. The paper evaluates prompt-derived and receiver-optimized payloads in auctions, reputation reporting, Werewolf, and blackjack, including heterogeneous model pairs. It also compares prompt-based interventions with replacing suspicious handoffs by their event-specific benign counterparts, and distinguishes local behavioral recovery from downstream task outcomes.

Critical strengths
  • The paper addresses an important monitoring gap when agents communicate outside the public transcript. Linking a latent handoff to the receiver's subsequent action is useful, and the distinction between detectability, behavioral influence, and successful coordination avoids interpreting every anomalous representation as a successful attack.

  • The evaluation covers four environments, four model families, and cross-family communication through benign-trained adapters. Including receiver-optimized payloads strengthens the evaluation beyond prompt-derived residuals. Reporting cases where detection remains possible despite little behavioral impact also helps characterize what the monitor actually measures.

Critical weaknesses
  • The practical defense relies on a benign reference for the same event that is not clearly available to the defender. The action score uses a control execution with the collusive handoff blocked, while the intervention replaces the suspect handoff with its corresponding benign counterpart. Access to a pool of benign traffic does not by itself provide this event-specific counterpart. Moreover, near-perfect recovery toward this reference is largely expected after restoring that input. Section 7 acknowledges the limitation, but practical detection and mitigation without this reference remain unestablished.

  • The operational benefit of the three-stage gate is insufficiently evaluated. Section 4 flags a handoff only when all three scores exceed their benign thresholds, but AUROC does not establish recall at that operating point. The combined scores also do not consistently improve over individual signals: in homogeneous reputation monitoring, L2 achieves 0.881 AUROC versus 0.754 for the sequential score and 0.806 for the continuous score. Fixed-threshold recall, false positives, and end-to-end harm reduction are needed to justify the chosen combination.

Questions for authors
  • What information allows the defender to obtain the event-specific benign handoff without knowing the attack construction? How do detection and intervention perform with an approximate reference built only from benign calibration data?
  • At the frozen benign-calibrated thresholds, what are the recall and false-positive rates of the three-stage gate versus the individual scores, and what downstream harm remains after applying the resulting interventions?
1 comment
AuthorsOpenReview ↗

Response to reviewer CMSV

Q1. What information allows the defender to obtain the event-specific benign handoff without knowing the attack construction? How do detection and intervention perform with an approximate reference built only from benign calibration data?

Ans: We would like to clarify that PATCH detection and intervention does not require any knowledge of attack construction. L1 and L3 detection monitors are based on neutral or benign latent distributions, and L2 measures the change (KL-divergence) and harmful shift in the receiver’s action distribution relative to the blocked handoff between the sender and the receiver. For harmful shift, L2 does not require attack-specific details but high-level knowledge of task-specific harmful actions such as low bids in the auction case study.

Intervention is performed only on the flagged/detected handoffs. Contextually matched-neutral replacement (reported intervention for white-box steering in the paper) replaces the flagged sender's message with a non-collusive objective such as “Choose independently from the legal bids; no bid target is assigned.” for the auction case study, which does not require any knowledge of the attack construction.

In addition to matched neutral replacement, we also evaluate two new practical interventions on sequential monitor for rebuttal: 1) blocking flagged handoffs and, 2) replacing flagged handoffs with benign vectors retrieved from the calibration pool using public context such as auction base price, receiver, previous price, and round in case of auction case study. Neither requires the attack/control pairing. Selective blocking restores the auction lowest-bid rate from 100% to the 23.2% no-handoff baseline. Calibration-based replacement reduces this rate to 28.8%. In reputation case study, both interventions prevent all 250 harmful reports.

Q2. At the frozen benign-calibrated thresholds, what are the recall and false-positive rates of the three-stage gate versus the individual scores, and what downstream harm remains after applying the resulting interventions?

Ans. The following new Tables R1 and R2 report the actual sequential AND gate at the frozen benign-calibrated 95th percentile threshold, separately from its AUROC ranking score.

For the new experiments, we evaluated PATCH's Sequential Monitor on Qwen3-0.6B for auction and reputation case studies. The detector was fitted on 2,000 benign events, with thresholds frozen on 500 separate benign calibration events. Each attack condition contains 250 held-out events.

Table R1. Detection and intervention against receiver-optimized latent attacks at frozen thresholds.

MetricAuctionReputation
Optimized attacks detected250/250250/250
Benign handoffs flagged0/27505/750
Measured attack-induced harm prevented100%100%
Benign actions changed by selective blocking0%0.7%
Benign actions changed by blocking all handoffs56.5%90.4%

Against the evaluated receiver-optimized attacks, PATCH detects 250/250 handoffs in each setting, with 0/2750 and 5/750 benign false-positive flags. Selective blocking prevents all measured attack-induced harm, with every attack event included in the denominator.

Table R2. Unnecessary interventions at frozen thresholds. All detectors retain 250/250 recall on the evaluated optimized attacks.

DetectorAuction (2750 benign)Reputation (750 benign)
L1: Representation anomaly1474270
L2: Action shift1948
L3: Sparse-feature novelty1,549277
PATCH*05

The individual signals (L1, L2, and L3) achieve the same recall on these optimized attacks but produce more unnecessary interventions. In reputation reporting, PATCH reduces false-positive interventions from 48 for L2 alone to 5, a 9.6-fold reduction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.