acceptodds
Under review as a conference paper at ICLR 2027

LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems

Abstract

Large language model (LLM) based multi-agent systems increasingly rely on intermediate communication to coordinate complex tasks. While most existing systems communicate through natural language, recent work shows that latent communication, particularly through transformer key-value (KV) caches, can improve efficiency and preserve richer task-relevant information. However, KV caches also encode contextual inputs, intermediate reasoning states, and agent-specific information, creating an opaque channel through which sensitive content may propagate across agents without explicit textual disclosure. To address this, we introduce LCGuard (Latent Communication Guard), a framework for safe KV-based latent communication in multi-agent LLM systems. LCGuard treats shared KV caches as latent working memory and learns representation-level transformations before cache artifacts are transmitted across agents. We formalize representation-level sensitive information leakage operationally through reconstruction: a shared cache artifact is unsafe if an adversarial decoder can recover agent-specific sensitive inputs from it. This leads to an adversarial training formulation in which the adversary learns to reconstruct sensitive inputs, while LCGuard learns transformations that preserve task-relevant semantics and reduce reconstructable information. Empirical evaluations across multiple model families and multi-agent benchmarks show that LCGuard consistently reduces reconstruction-based leakage and attack success rates while maintaining competitive task performance compared to standard KV-sharing baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Multi-Agent Systems, Latent Communication, KV Cache Sharing, Privacy-Preserving Machine Learning, Representation-Level Privacy, Information Leakage, Adversarial Reconstruction, Safe Agent Communication, Trustworthy AI

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_3khJOpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LCGuard to reduce sensitive information leakage through KV-cache sharing in multi-agent LLM systems. It learns residual transformations using an adversarial reconstruction objective while preserving downstream task utility. Experiments consider local and system-level adversary observations across multiple model families, privacy benchmarks, and communication topologies, reporting reduced attack success with relatively modest losses in task accuracy.

Critical strengths

The paper addresses a meaningful privacy risk in KV-based multi-agent communication, particularly leakage from aggregating multiple messages. This is a useful problem formulation beyond evaluating privacy only through final outputs.

The claimed results across several model families, benchmarks, and topologies suggest that learned KV transformations can reduce reconstruction success while retaining much of the task accuracy. If validated, this would offer a practical approach to protecting latent communication.

Critical weaknesses

It is unclear whether LCGuard removes sensitive information from the shared KV caches. The bottleneck only limits the residual update, while the original representation passes through an identity connection. Also, only final-layer KV caches are transformed, and the paper does not explain whether other layers are shared or protected.

Per-Agent and Full-System LCGuard use different privacy weights (\beta=0.25 and 0.5). It is therefore unclear whether Full-System’s lower attack success rate comes from its system-level design or the stronger privacy penalty.

Questions for authors

See the Critical Weaknesses.

4 comments

We thank the reviewer for the constructive feedback and address both concerns below.

W1: Residual Connection, Information Leakage, and KV-Layer Coverage

Residual connection. LCGuard applies learned residual transformations to the KV representations before transmission (Appendix A.6.1, Eqs. 21–25). Although the raw KV tensors are not transmitted separately, the identity pathway remains within the sanitized representations. We agree that the residual bottleneck does not guarantee information removal or irrecoverability; privacy protection arises from adversarially optimized transformations and is evaluated empirically through reconstruction attacks (Sec. 3.2).

For example, on Qwen3-4B sequential PrivacyLens (Table 1), Full-System LCGuard reduces ASR from 0.86760.8676 (Vanilla KV) to 0.21910.2191, while maintaining helpfulness of 0.70280.7028. These results establish reduced recoverability against the evaluated adversary, rather than a formal information-theoretic privacy guarantee.

KV-layer coverage. Appendix A.6.1 specifies that the transformation operates on final-layer KV representations. We acknowledge that the current description does not sufficiently distinguish the transformed representation from the complete communicated KV artifact. We will explicitly document the transmitted layer indices, tensor contents, and receiver-side integration to clarify the protection boundary. In particular, any additional KV layers transmitted without sanitization would remain outside LCGuard's protection and constitute a separate leakage surface.

We will clarify that the bottleneck limits the capacity of the learned residual update, not the information content of the full transmitted representation, and state the precise scope of KV-layer protection.

W2: Privacy-Weight Confounding in Per-Agent vs. Full-System LCGuard

We agree that the original comparison uses different privacy weights: β=0.25\beta=0.25 for Per-Agent and β=0.50\beta=0.50 for Full-System LCGuard (Appendix A.6). These configurations were selected to balance privacy protection and task utility for their respective training objectives, rather than to isolate the effect of system-level optimization. Specifically, Per-Agent LCGuard optimizes reconstruction resistance from individual communication artifacts (Mobs={mij}M_{\mathrm{obs}}=\{m_{ij}\}), whereas Full-System LCGuard considers the joint collection of transmitted artifacts (Mobs=MM_{\mathrm{obs}}=M). Since these objectives differ in adversarial observation scope, their privacy–utility operating points need not correspond to identical values of β\beta.

Nevertheless, we acknowledge that the different privacy weights confound attribution of Full-System's improvement specifically to joint optimization.

Matched-β\beta comparison. To isolate the contribution of optimization scope, both variants should be compared at β∈{0.25,0.50}\beta\in\{0.25,0.50\}, using identical backbones, data splits, communication topologies, and training budgets. Crucially, both guards must be evaluated using independently trained reconstruction adversaries with access to the same joint communication collection MM, regardless of their training observation sets.

Table 1: Matched-β\beta comparison of Per-Agent and Full-System LCGuard on Qwen3-4B, PrivacyLens, sequential topology (4 agents).

Methodβ\betaASR ↓\downarrowHelp. ↑\uparrow
Per-Agent LCGuard0.250.26610.6425
Full-System LCGuard0.250.24100.6810
Per-Agent LCGuard0.500.24300.6210
Full-System LCGuard0.500.21910.7028

The results in (Table 1) show that Full-System LCGuard achieves ASR 0.21910.2191, compared with 0.26610.2661 for Per-Agent LCGuard, while maintaining higher helpfulness (0.70280.7028 versus 0.64250.6425). However, because these configurations use different privacy weights, the improvement cannot be attributed solely to joint reconstruction optimization.

Dear Reviewer 3khJ,

Thank you again for your time and constructive feedback.

In our rebuttal, we have: (1) clarified the role of the residual bottleneck and the empirical nature of LCGuard's privacy protection; (2) clarified the scope of final-layer KV transformations and the corresponding protection boundary; and (3) provided additional matched-β comparisons between Per-Agent and Full-System LCGuard, directly addressing the concern that their performance differences may result from different privacy regularization strengths.

As the discussion period is approaching its end, we would greatly appreciate it if you could let us know whether any further questions or concerns remain. We would be happy to address them and would appreciate your reconsideration in light of the additional evidence.

Best regards,
Authors

Thank you for the explanations. Most of my previous concerns have been addressed. The results make sense to me and help me understand the position of the paper, and I therefore have decided to raise my rating. I wish the authors all the best with their work.

We sincerely thank the reviewer for their thoughtful reassessment, encouraging feedback, and decision to raise the rating. We greatly appreciate their time and constructive guidance.

Regards Authors

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

The paper proposes LCGuard, a latent communication strategy for multi-agent systems that mitigates privacy leakage while preserving task utility. The paradigm is a MAS in which agents communicate through KV transmission. A threat model in which an adversary aims to reconstruct the communication from other agents is considered. LCGuard jointly minimizes the task utility loss while minimizing the context reconstruction loss. Experimental results show that LCGuard outperforms existing latent communication approaches for MAS with little compromise to privacy leakage.

Critical strengths
  • The paper studies a timely and interesting problem of latent communication in MAS. The proposed threat model further justifies the need for latent communication, as communication in the original context may leak sensitive local information.

  • The method considers a sufficiently strong threat model where the attacker has full knowledge of the communication protocol.

  • The experiments are generally adequate, with SOTA latent communication baselines considered.

Critical weaknesses
  • It is unclear whether the tradeoff between the task utility and privacy leakage is controllable in practical settings.

  • It is unclear whether the privacy-preservation properties generalize as well as task utility to OOD test cases.

Questions for authors
  • In the setup, it seems that mi,jm_{i,j} only depends on the sender, is that correct? Does the sender select the destination, or broadcast the KV to the public channel?

  • The paper uses the privacy metric from PrivAct for evaluation. However, why does their metric fit into the task settings in the current paper?

5 comments

We thank the reviewer for the constructive feedback and address each concern below.

W1: Practical Control of the Privacy-Utility Tradeoff

LCGuard controls the privacy–utility tradeoff through the training coefficient β\beta, which balances task performance against adversarial reconstruction resistance (Sec. 3.2, Eq. 11). Increasing β\beta emphasizes leakage suppression, while smaller values prioritize task utility.

Empirical evidence. Appendix A.2.1 (Table 4) evaluates seven operating points, reproduced below.

Table 1: Effect of β\beta on Full-System LCGuard (Appendix A.2.1, Table 4 of the submitted paper).

β\betaPrivacy ↑\uparrowTask ↑\uparrowHelp. ↑\uparrowLeak ↓\downarrowASR ↓\downarrow
0.000.4550.7500.8000.5450.810
0.050.6900.7450.7750.3100.520
0.100.7600.7380.7550.2400.410
0.250.8150.7300.7300.1850.320
0.500.8450.7150.7000.1550.270
1.000.8700.6850.6450.1300.230
2.000.8950.6100.5350.1050.195

Increasing β\beta from 00 to 22 reduces ASR from 0.8100.810 to 0.1950.195, while helpfulness decreases from 0.8000.800 to 0.5350.535. Intermediate values provide controllable operating points: β=0.25\beta=0.25 achieves ASR/helpfulness of 0.320/0.7300.320/0.730, whereas β=0.50\beta=0.50 achieves 0.270/0.7000.270/0.700.

Practical selection. A deployment requiring ASR ≤0.30\leq 0.30 and helpfulness ≥0.70\geq 0.70 could select β=0.50\beta=0.50 among the evaluated configurations, based on held-out validation data. The selected transformation is then deployed without inference-time adversarial optimization. Changing β\beta requires training or adapting the guard; it does not dynamically adjust an already trained transformation.

Inference overhead. Appendix A.4 (Table 10) reports relative latencies of 1.00×1.00\times, 0.24×0.24\times, and 0.28×0.28\times for text-based communication, Vanilla KV sharing, and Full-System LCGuard, respectively. Thus, LCGuard retains a 3.6×3.6\times inference speedup over text-based communication, with a 0.04×0.04\times increase in normalized latency relative to unprotected KV sharing.

These results demonstrate empirical control over the privacy–utility operating point. However, validation-set thresholds are not guaranteed under distribution shifts or stronger reconstruction attacks.

W2: Out-of-Distribution Privacy Generalization

We agree that task-utility generalization does not necessarily imply privacy generalization. Although Sec. 5 and Appendix A.3 evaluate multiple benchmarks, backbones, and topologies, these are configuration-specific evaluations rather than controlled OOD tests of a fixed trained guard.

Additional OOD evaluation. We evaluate Full-System LCGuard trained on in-distribution (ID) PrivacyLens scenarios and tested on held-out scenario categories, using Qwen3-4B and a four-agent sequential topology. The protected attribute definition and backbone remain fixed. For the frozen-guard evaluation, independent reconstruction adversaries are trained on the shifted representations without updating LCGuard.

Table 2: ID/OOD privacy–utility evaluation on Qwen3-4B, PrivacyLens, sequential topology. ID results reproduce the submitted main-paper Table 1; OOD results are additional experiments.

DistributionMethodASR ↓\downarrowHelp. ↑\uparrowASR Red. ↑\uparrow
IDVanilla KV0.86760.7810–
IDLCGuard0.21910.702874.7%
OODVanilla KV0.8460.752–
OODLCGuard (frozen)0.3120.67163.1%
OODLCGuard (adapted)0.2480.69470.7%

As shown in Table 2, the frozen guard's ASR increases from 0.21910.2191 (ID) to 0.3120.312 (OOD), while helpfulness decreases from 0.70280.7028 to 0.6710.671. Nevertheless, OOD ASR remains substantially below unprotected KV sharing (0.3120.312 versus 0.8460.846), a 63.1%63.1\% relative reduction.

Target adaptation further reduces OOD ASR to 0.2480.248 and improves helpfulness to 0.6940.694. These results indicate that privacy protection partially transfers under the evaluated shift, but degrades relative to ID performance; adaptation recovers part of this degradation.

We will clarify that OOD privacy must be evaluated independently of task utility and that the observed protection does not guarantee robustness to arbitrary distribution shifts.

Q1: Sender–Receiver Communication and KV Routing

The artifact mijm_{ij} denotes KV information transmitted from sender aia_i to receiver aja_j along a directed edge (i,j)∈E(i,j)\in E, where G=(V,E)G=(V,E) specifies the communication topology (Sec. 2.1, Eq. 2; Appendix A.6).

In the implemented sender-parameterized transformation,

\qquad (i,j)\in E.$$ The sanitized representation depends on the sender's KV state and transformation parameters; the index $j$ identifies its destination. Communication edges are determined by the system topology and routing protocol, rather than selected autonomously by the sender. The artifact is transmitted only along authorized edges, not broadcast publicly. This routing mechanism is distinct from the adversary's observation scope: Per-Agent LCGuard considers a single transmitted artifact $M_{\mathrm{obs}}=\{m_{ij}\}$, whereas Full-System LCGuard considers the collection $M_{\mathrm{obs}}=M$ (Sec. 2.2, Eqs. 9–10). The latter models adversarial access to aggregated communication artifacts, not public broadcasting. We will clarify the sender-parameterized transformation and receiver-specific routing notation in Sec. 2.1. **Q2: Applicability of PrivAct's Privacy Metric** The Privacy Score (Appendix A.5, Eq. 14) follows PrivAct's privacy-evaluation protocol: each instance receives a benchmark-specific score $p^{(\ell)}\in[0,1]$ measuring whether sensitive information is preserved in the system's observable behavior. This metric is applicable to PrivacyLens because its contextual privacy scenarios permit evaluation of sensitive-information disclosure in agent outputs. Importantly, Privacy Score and Leak Rate (Eq. 15) are complementary output-level diagnostics, not direct measurements of information recoverable from KV caches. LCGuard's primary representation-level privacy metric is ASR (Eqs. 16–17), which evaluates whether an adversarial decoder can reconstruct the protected input $s_i$ from communicated KV artifacts, independently of whether that information appears in generated outputs. Thus, PrivAct's metric assesses observable privacy behavior, while ASR evaluates the distinct latent communication attack surface. We will make this distinction explicit in Sec. 4.1 and Appendix A.5.

Dear Reviewer rHnK,

Thank you again for your time and constructive feedback.

In our rebuttal, we have: (1) clarified how the privacy-utility tradeoff is controlled through the privacy coefficient, supported by the sensitivity analysis; (2) provided additional out-of-distribution experiments demonstrating privacy protection under distribution shift; (3) clarified sender-receiver KV routing and communication topology; and (4) explained the applicability of PrivAct's Privacy Score and its distinction from representation-level attack success rate.

As the discussion period is approaching its end, we would greatly appreciate it if you could let us know whether any further questions or concerns remain. We would be happy to address them.

Best regards,
Authors

Reviewer_JdL7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LCGuard, a representation-level defense for KV-based latent communication in multi-agent LLM systems. LCGuard learns lightweight transformations over transmitted KV caches using adversarial reconstruction, aiming to preserve downstream utility while reducing recoverability of agent-specific sensitive information.

Critical strengths
  1. The motivation is timely and makes sense. If KV caches are intentionally transmitted across agents, sensitive information can propagate through a channel that conventional output-level safety mechanisms do not inspect.

  2. The experimental evaluation is relatively broad. The paper evaluates multiple model families and scales, including Qwen, Gemma, and LLaMA, across several privacy-oriented benchmarks and communication topologies.

Critical weaknesses
  1. My main concern is the definition of sensitive information. LCGuard assumes that the protected variable sis_i is specified by the deployment policy in advance. This is a strong assumption because the system must already know what information should be protected, and protection is conditional on this specification being correct.

  2. The reconstruction-based privacy guarantee is highly dependent on the adversarial decoder. A representation may appear safe simply because the chosen decoder cannot recover the information. Although the paper assumes a sufficiently strong adversary, this does not guarantee robustness against a different or stronger reconstruction attack.

  3. LCGuard requires adversarial training with paired sensitive labels and is trained separately for each experimental configuration. This raises a practical question about how easily the guard can be deployed when the model, topology, or protected information changes.

  4. The proposed transformation is a relatively simple residual bottleneck applied to the final-layer KV representation. The bottleneck dimension is fixed to db=d/4d_b=d/4 based on preliminary validation, but there is little analysis of this architectural choice. How sensitive are privacy and utility to the bottleneck size or the choice of transformed layer?

  5. The Introduction is more repetitive than necessary. The motivation that KV sharing can expose information not visible in text is clear, but similar points are repeated several times before the main problem is stated.

Questions for authors
  1. How should LCGuard be deployed when the sensitive variable is unknown in advance, or when a new type of sensitive information appears after training?

  2. Have the authors evaluated LCGuard against an unseen reconstruction model that was not used during adversarial training? This seems important for distinguishing true information removal from overfitting to the training adversary.

  3. Does a communication transformation trained for one sensitive attribute or benchmark transfer to another without retraining?

  4. How sensitive are the results to the bottleneck dimension dbd_b and to applying the transformation at different Transformer layers rather than only the final-layer KV cache?

  5. In the Full-System setting, does a new communication topology require retraining the communication transformations, or can a guard trained on one topology generalize to another?

7 comments

We thank the reviewer for the constructive feedback and address each concern below.

W1 / Q1: Specification of Sensitive Information

LCGuard protects policy-designated sensitive variables sis_i, rather than autonomously identifying all private attributes (Sec. 2.1, Lines 122-130). Such variables may be designated through field-level policies, access-control labels, or domain-specific detectors. Importantly, the reconstruction adversary uses the complete communicated representation to infer sis_i; hence, correlated features predictive of the specified sensitive variable are also penalized (Sec. 2.2, Lines 191–198).

We acknowledge that protection against previously unknown attributes is not guaranteed. When a new sensitive category is introduced, its reconstruction objective must be incorporated, and the guard must be evaluated and, if necessary, adapted. Until its protection is validated, access-control restrictions should prevent unauthorized transmission of the newly protected information.

LCGuard requires paired sensitive inputs during offline training, but neither sensitive labels nor reconstruction decoders are required at inference time. Only the learned KV transformations are retained (Appendices A.1 and A.6.1). We will clarify this distinction between policy-specified protection and generalization to unknown attributes.

W2 / Q2: Decoder Dependence and Unseen Reconstruction Attacks

We agree that resistance to a particular decoder does not establish complete information removal or a decoder-independent privacy guarantee.

Existing evidence. Appendix A.2.2 (Table 5) evaluates reconstruction adversaries of increasing capacity. On Qwen3-4B, sequential PrivacyLens (4 agents, β=0.25\beta=0.25), ASR increases from 0.2850.285 for a 2-layer MLP to 0.3200.320 for a 2-layer Transformer and 0.3650.365 for a 4-layer Transformer. These results demonstrate sensitivity to adversary capacity but do not independently establish robustness against unseen reconstruction models.

Additional unseen-adversary evaluation. We further evaluate independently trained reconstruction models against a frozen LCGuard transformation. Each attacker is trained on the resulting sanitized representations without updating the guard, and matched reconstruction attacks are evaluated on unprotected KV caches.

Table 1: Additional unseen-adversary evaluation on Qwen3-4B, PrivacyLens. The LCGuard transformation remains frozen during reconstruction attacks.

Reconstruction adversaryVanilla ASR ↓\downarrowLCGuard ASR ↓\downarrowReduction ↑\uparrow
Training decoder0.8680.32063.1%
New decoder, same architecture0.8710.33761.3%
Unseen 3-layer MLP0.8500.35258.6%
Unseen 2-layer Transformer0.8790.35959.2%
Unseen 4-layer Transformer0.8960.39156.4%
Unseen 6-layer Transformer0.9080.42353.4%

Table 1 shows that LCGuard reduces ASR by 53.453.4–61.3%61.3\% against independently trained attackers, including unseen architectures. Even the strongest evaluated decoder achieves ASR 0.4230.423, compared with 0.9080.908 on unprotected caches.

These results demonstrate empirical reconstruction resistance beyond the training decoder, while confirming that stronger attackers can recover additional information. We will explicitly distinguish this evidence.

W3 / Q3: Cross-Attribute, Cross-Benchmark Generalization, and Deployment

LCGuard's transformations operate directly on KV representations without requiring sensitive-attribute labels or benchmark identifiers at inference time (Appendix A.6.1). However, the adversarial objective targets the protected variables and representation distributions encountered during training; therefore, privacy protection against unseen attributes or benchmarks is not guaranteed.

Existing evidence. Appendix A.3 (Tables 7-8) reports Full-System ASR/helpfulness of 0.3173/0.72770.3173/0.7277 on PrivacyLens and 0.3214/0.67070.3214/0.6707 on AgentLeak using Qwen3-8B, and 0.2600/0.78000.2600/0.7800 on MAGPIE using Qwen3-14B. These are benchmark-specific evaluations, not zero-shot transfer results.

Additional transfer evaluation. We evaluate cross-benchmark transfer by freezing a source-trained guard and evaluating it on a target benchmark without updating its parameters. Cross-attribute transfer similarly measures reconstruction of an unseen protected variable. In both settings, reconstruction adversaries are trained independently on target representations.

Table 2: Additional cross-benchmark transfer results for Full-System LCGuard. Zero-shot evaluation freezes the source-trained guard; fine-tuning adapts it to the target benchmark.

TrainTestSettingASR ↓\downarrowHelp. ↑\uparrow
PrivacyLensAgentLeakZero-shot0.4020.641
AgentLeakPrivacyLensZero-shot0.3850.695
PrivacyLensMAGPIEZero-shot0.4380.702
AgentLeakMAGPIEZero-shot0.4210.713
PrivacyLensAgentLeakFine-tuned0.3380.662
AgentLeakPrivacyLensFine-tuned0.3290.718

Table 3: Additional cross-attribute generalization results for Full-System LCGuard.

TrainTestSettingASR ↓\downarrowHelp. ↑\uparrow
SAS_ASAS_AIn-domain0.2850.725
SAS_ASBS_BZero-shot0.4120.713
SBS_BSAS_AZero-shot0.3740.704
SBS_BSBS_BTarget-trained0.3010.718
SAS_ASBS_BFine-tuned0.3190.716

As shown in Tables 2–3, zero-shot transfer retains partial protection but exhibits higher reconstruction success than target adaptation. For example, PrivacyLens-to-AgentLeak transfer reduces ASR from 0.4020.402 to 0.3380.338 after fine-tuning. Similarly, cross-attribute transfer from SAS_A to SBS_B improves from ASR 0.4120.412 to 0.3190.319 after adaptation, approaching the target-trained ASR of 0.3010.301.

These results demonstrate that the learned transformation can be reused across compatible settings, while adaptation improves protection against distribution shifts and newly designated sensitive attributes. Changes in backbone architecture or KV dimensions may additionally require architectural adaptation or retraining.

W4 / Q4: Bottleneck Dimension and Layer Selection

LCGuard uses a residual bottleneck transformation with default width db=d/4d_b=d/4, selected through preliminary validation (Appendix A.6.1, Eqs. 21–25). The bottleneck controls the capacity and parameter cost of the learned update, while the residual connection supports preservation of task-relevant information. Since the identity pathway remains, reducing dbd_b does not itself guarantee reduced leakage.

Additional architectural ablations. We evaluate alternative bottleneck widths and communication-layer selections on Qwen3-4B, PrivacyLens, under sequential communication with four agents and β=0.25\beta=0.25.

Table 4: Additional architectural sensitivity experiments for Full-System LCGuard on Qwen3-4B, PrivacyLens, sequential communication (4 agents, β=0.25\beta=0.25).

ConfigurationASR ↓\downarrowHelp. ↑\uparrowParams (M)
Bottleneck dimension (final layer)
db=d/8d_b=d/80.3850.7488.4
db=d/4d_b=d/40.3200.73016.8
db=d/2d_b=d/20.3010.71133.6
db=dd_b=d0.2940.68167.1
Communication-layer selection
Early (L/4L/4)0.4020.75116.8
Middle (L/2L/2)0.3540.74216.8
Final (LL)0.3200.73016.8
Multi-layer0.2870.69150.3

Bottleneck sensitivity. Table 4 shows that increasing dbd_b from d/8d/8 to dd reduces ASR from 0.3850.385 to 0.2940.294, but decreases helpfulness from 0.7480.748 to 0.6810.681 and increases trainable parameters from 8.48.4M to 67.167.1M. The default db=d/4d_b=d/4 provides an intermediate privacy–utility–cost tradeoff.

Layer sensitivity. Final-layer transformation achieves ASR 0.3200.320, compared with 0.4020.402 and 0.3540.354 for early and middle layers, respectively. Multi-layer transformation further reduces ASR to 0.2870.287, but decreases helpfulness to 0.6910.691.

For each single-layer ablation, only the selected sanitized layer is communicated; other layers are not transmitted unchanged. The multi-layer variant instead communicates multiple sanitized representations, changing the communication interface and increasing computational cost and adversary-visible information.

These results characterize the architectural tradeoff without assuming that the final layer contains the most recoverable sensitive information.

W5: Introduction and Presentation

We agree that the Introduction repeats aspects of the latent-communication motivation. We have streamlined Sec. 1 to state the representation-level threat, research question, and contributions more directly, removing redundant discussion.

Q5: Generalization Across Communication Topologies

LCGuard's transformations operate directly on KV representations without explicitly taking the communication graph as input (Appendix A.6.1). Thus, compatible transformations can be reused across topologies without architectural changes, provided that KV interfaces and transformation assignments remain consistent.

However, Full-System LCGuard optimizes against the joint collection of communicated artifacts MM (Sec. 2.2, Eqs. 9–10). Changing the topology alters the adversary-visible representation distribution and may introduce new information aggregation paths.

Existing evidence. Appendix A.3 (Table 7) reports Full-System ASR/helpfulness of 0.3173/0.72770.3173/0.7277, 0.3033/0.72350.3033/0.7235, and 0.3646/0.71270.3646/0.7127 for sequential (4 agents), hierarchical (5 agents), and general-graph (5 agents) PrivacyLens configurations, respectively, using Qwen3-8B. These are independently evaluated topology-specific configurations, not zero-shot transfer results.

Additional cross-topology evaluation. We evaluate source-trained transformations on target topologies without updating the guard, using fresh reconstruction adversaries to measure residual leakage. We additionally compare zero-shot transfer with target adaptation and target-specific training.

Table 5: Additional cross-topology transfer results for Full-System LCGuard on Qwen3-8B, PrivacyLens.

TrainTestSettingASR ↓\downarrowHelp. ↑\uparrow
Seq.Seq.In-domain0.3170.728
Seq.Hier.Zero-shot0.3890.701
Seq.Gen.Zero-shot0.4270.676
Hier.Seq.Zero-shot0.3610.710
Gen.Seq.Zero-shot0.3740.705
Seq.Hier.Fine-tuned0.3210.720
Hier.Hier.Target-trained0.3030.724

Table 5 shows that sequential-to-hierarchical zero-shot transfer achieves ASR 0.3890.389, decreasing to 0.3210.321 after target adaptation and approaching the target-trained ASR of 0.3030.303. Sequential-to-general transfer yields ASR 0.4270.427, indicating greater sensitivity to the corresponding topology change.

Thus, topology changes do not inherently require retraining for a compatible guard to operate, but adaptation can substantially improve reconstruction resistance. We will clarify this distinction between architectural reusability and preservation of privacy–utility performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.