acceptodds
Under review as a conference paper at ICLR 2027

RouteTrace: Membership Inference from Routed Update Traces in Mixture-of-Experts Models

Abstract

Router updates in a sparse mixture-of-experts (MoE) model can reveal whether a record was used for post-training. We introduce ROUTETRACE, a white-box membership audit of paired base and post-trained checkpoints. ROUTETRACE measures how a candidate’s negative router gradient aligns with the net router update, both across layers and along the candidate’s route. The candidate’s token routes identify router rows associated with selected experts, allowing this alignment to be compared on selected and unselected rows. With three matched shadow checkpoints for each model and dataset, its learned scorers reach mean AUCs of .948 on Natural Questions and .862 on Dolly across Qwen, DeepSeek-MoE, and OLMoE. On Natural Questions, global depth features, which pool evidence across all router rows, preserve most of the full scorer’s ranking accuracy, while alignment on selected rows separates members more strongly than equal-sized next-ranked and random unselected controls

Primary area:
foundation or frontier models, including LLMs
Keywords:
MoE Models, Membership Inference Attack, LLM Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pFoTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RouteTrace, a white-box membership inference attack against Mixture-of-Experts models. RouteTrace uses the alignment between router gradients and parameter updates observed between base and post-trained checkpoints. The method combines global alignment, layerwise information, and route-conditioned features to infer whether a candidate record was included in the post-training dataset.

The authors evaluate RouteTrace on three MoE model families and two datasets. The method shows better performance than some output-based MIA baselines, although LiRA and RMIA achieve higher AUC in the reported Qwen experiments when sufficient reference models are available.

Critical strengths

1. Transparent reporting and analysis of negative results. The authors transparently report the limitations of their proposed approach. RouteTrace has lower performance compared to LiRA and RMIA in the evaluated reference-based setting.

2. Investigation of membership leakage through MoE router updates. The paper investigates an important privacy concern in Mixture of Experts models by examining whether router parameter updates reveal membership information.

Critical weaknesses

1. Limited training scale and insufficient analysis of its impact on membership inference The experimental evaluation is conducted under small-scale fine-tuning settings. Specifically, Natural Questions uses 4,096 training samples and 256 optimizer steps, while Dolly uses only 1,024 samples and 64 optimizer steps. Both settings involve only a single epoch of fine-tuning. Since RouteTrace directly exploits the alignment between candidate gradients and router parameter updates, it is unclear whether the reported attack performance would remain effective under larger-scale fine-tuning settings.

Although Appendix H.4 includes a limited exposure comparison, it does not systematically show the effects of dataset size and training duration under the primary fine-tuning configuration.

2. Restricted fine-tuning setup and potential dependence on router updates. The primary experimental setup is updating the router matrices and rank-8 LoRA adapters on the attention Q/V projections while freezing the expert parameters. This setting is relatively restrictive and may favor RouteTrace, since the proposed attack directly exploits router parameter updates. It is unclear whether the reported effectiveness generalizes to other fine-tuning configurations, particularly those that also update expert parameters.

3. Incomplete comparison with strong reference-based MIA baselines.

The evaluation of LiRA and RMIA is restricted to two Qwen targets on Natural Questions. However, these methods achieve higher AUC than RouteTrace when sufficient reference checkpoints are available.

4. Insufficient discussion of shared experts. The evaluated MoE architectures differ in their use of shared experts. In particular, Qwen and DeepSeek-MoE incorporate shared experts alongside routed experts.

However, the paper does not clearly discuss the implications of shared experts for the proposed membership inference mechanism. Since shared experts are activated independently of the top-k routing decision, they may influence the representations and gradient signals used by RouteTrace.

5. Missing sensitivity analysis of LoRA rank. The primary experiments use a fixed LoRA rank of 8 for the attention Q/V adapters. However, the paper does not investigate how different LoRA ranks affect membership inference performance.

Since LoRA rank controls the adaptation capacity of the attention modules, varying this parameter may change the learned representations and the corresponding router gradients used by RouteTrace.

Questions for authors
  • How does RouteTrace's performance change as the fine-tuning dataset size increases while keeping the model architecture and training configuration fixed?
  • How sensitive is the attack to the number of fine-tuning steps or epochs?
  • How does RouteTrace's performance vary across different LoRA ranks?
  • Why are LiRA and RMIA evaluated only on two Qwen Natural Questions targets? Extending these evaluations to the remaining model families and datasets would provide a more comprehensive assessment of the results.
  • What motivates the choice of updating router matrices and Q/V adapters while freezing all expert parameters?
  • How does RouteTrace perform under alternative fine-tuning configurations, such as expert fine-tuning or full-model fine-tuning?
  • How are shared experts and their associated components treated during fine-tuning and RouteTrace feature extraction?
2 comments
AuthorsOpenReview ↗

Response on experimental scope

Thank you for identifying these scope limitations.

Training scale. The main experiments use one epoch: 4,096 training records for NQ and 1,024 for Dolly. The existing exposure contrast is not a systematic dataset-size or training-duration study. We did not test larger datasets or more epochs.

Fine-tuning configuration. The experiments update routers and rank-8 Q/V LoRA adapters while freezing experts. We did not evaluate expert fine-tuning, full-model fine-tuning, or other LoRA ranks. The current evidence is therefore limited to router-changing post-training.

LiRA/RMIA. The strict reference-based comparison covers two Qwen NQ targets. With 16 references, LiRA and RMIA have higher AUC than RouteTrace, which uses three shadows. We did not build equivalent reference banks for the remaining models and datasets.

Shared experts. Shared experts are not selected by the top-k router and therefore do not enter the selected/boundary row sets. They can still affect hidden representations, the candidate loss, and thus router gradients. The Qwen and DeepSeek-MoE results include this indirect effect, but we did not isolate shared experts in a separate ablation.

Thank you for your rebuttal. My questions about the fine-tuning setup and LiRA/RMIA were requests for further evaluation. Testing a setting where both routers and experts are updated would help assess whether the findings depend on freezing the experts. Extending the LiRA/RMIA comparison beyond the two Qwen NQ targets would also help assess RouteTrace’s relative performance, taking the different reference and shadow-model budgets into account.

The explanation of shared experts is helpful. Since they can still affect router gradients indirectly, a controlled analysis of their impact on attack performance would strengthen the explanation of the mechanism.

My main concerns remain unresolved. I am therefore maintaining my original rating.

Reviewer_oF5wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes RouteTrace, a white-box membership inference attack for Mixture-of-Experts (MoE) models. It measures the alignment between candidate gradients and router parameter updates, using both global and route-specific features to identify training members. Experiments on three MoE models and two datasets show strong performance against loss-based baselines, with additional analyses of router-specific membership leakage. I apprecite the comprehensive experiment, but I still have some concerns on the novelty and results.

Critical strengths
  1. Comprehensive Experiments. The paper evaluates three MoE models and two datasets, with comparisons against multiple MIA baselines.

  2. Detailed Analysis. The paper provides extensive ablation studies and additional experiments to investigate membership leakage in MoE models.

Critical weaknesses
  1. Poor Presentation. The paper is difficult to read and contains excessive defensive explanations that interrupt the main discussion. The overall writing and organization need substantial improvement.

  2. Limited Novelty. The core gradient–update alignment idea has been explored in prior MIA studies. The paper does not clearly explain its methodological novelty or sufficiently distinguish itself from existing work.

  3. Limited Benefits of the Proposed Design. The ablation study shows that the full RouteTrace only improves AUC by 0.002–0.010 over the simpler Global Depth baseline. This raises questions about the necessity and effectiveness of the proposed route-conditioned features.

  4. Unclear Code Availability. The paper does not provide a clear link to the source code or a code release plan, making it difficult to verify and reproduce the experimental results.

Questions for authors
  1. Novelty. Prior studies (e.g., Li et al., ICLR 2023) have already used cosine similarity between candidate gradients and model parameter updates for MIA. This paper follows the same core idea, applying it to MoE routers and introducing route-conditioned features. Could the authors clarify the key differences from prior work and explain what limitations of existing methods their design addresses?

Effective passive membership inference attacks in federated learning against overparameterized models Li et al., ICLR 2023

  1. Method Design. RouteTrace uses a hybrid state (post-trained Q/V + base router) to compute gradients and identify selected experts, while the parameter updates come from the differences between the base and post-trained routers. If expert routing changes during post-training, can these selected experts accurately represent the routing paths used during training? Could the authors explain the rationale behind this design and its impact on attack performance?

  2. Ablation Study. Appendix H.1 shows that Full RouteTrace improves AUC by only 0.002–0.010 over the simpler Global Depth baseline. Given their similar performance, are the additional route-conditioned features really necessary?

  3. Practical Applicability. Appendix G shows that freezing the router makes RouteTrace ineffective, while task performance slightly improves across all three models. This suggests that updating the router may not be necessary in the evaluated setting. Could the authors explain why router updates are needed in practical MoE fine-tuning and provide further evidence of the relevance of this attack setting?

2 comments
AuthorsOpenReview ↗

Response on novelty and feature ablations

Thank you for the comments.

Novelty. We do not present gradient alignment itself as new. RouteTrace uses token-level expert selection to compare selected router rows with equal-size boundary and random rows, thereby localizing where the membership signal is concentrated.

Global Depth ablation. We agree that Global Depth retains most of the ranking performance. On the three Qwen targets, Full Trace improves over Global Depth by only .0003, .0008, and .0051, and only one paired interval excludes zero. The route-conditioned features primarily localize the signal; they do not provide a large AUC gain. We will revise the corresponding presentation.

Router freezing. RouteTrace applies to router-changing post-training. Q/V-only training makes the router update zero and RouteTrace constant. This is an applicability boundary, not a general privacy defense, and we do not claim that practical fine-tuning must update routers.

Code. The current submission does not provide a clear anonymous code link. We agree that this is a reproducibility omission.

Thanks for your rebuttal. However, this rebuttal is too short to fully answer my questions. I will keep my rate and will not support the acceptance of this paper.

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

This paper proposes a white-box membership inference attack (MIA) method tailored for sparse MoEs by leveraging paired base and post-trained checkpoints to detect training data inclusion. Its basic idea is to compute the cosine similarity between a candidate record’s negative router gradient (θ~\tilde{\theta}) and the net router update (ΔR=Rp−Rb\Delta R = R_p - R_b) across layers, contrasting selected expert router rows (SltS_{lt}) against boundary (BltB_{lt}) and random unselected rows. It was tested across three MoE families on natural questions and Dolly datasets showing that high-ranking AUC compared to output-based baselines, though trailing reference-based attacks like LiRA and RMIA when robust reference budgets are available.

Critical strengths

The paper attempts to tackle an important problem in AI privacy domain, namely white-box membership inference attack and evaluates the proposed method against cutting-edge open-source MoE architectures and some datasets, providing a relatively thorough breakdown across different model scales. Additionally, it extensively analyzes the impact of different analysis states (B,H,PB, H, P), parameter blocks (router vs. Q/V adapters), and feature groupings (global depth vs. conditional features). The paper also compares against a wide array of output-based baselines, loss profiles, Min-K%++, and EZ-MIA, which is definitely a strength.

Critical weaknesses
  1. Weak Problem Motivation and Generalizability: It is unclear why white-box membership inference should be restricted exclusively to MoE models. To the best of my knowledge, gradient-to-update alignment is a generalized technique applicable to any neural network tensor, and the paper provides no proper justification or motivation for why dense models cannot undergo the exact same treatment, rendering the method a siloed trick rather than a generalizable privacy finding. Furthermore, the core premise of requiring white-box access to both a base model and its exact post-trained checkpoint with perfectly matched, aligned router rows feels artificial, as auditors in practical deployment scenarios (e.g. commercial APIs or standard open-weight releases) almost never possess the precise pre-trained base model alongside the post-trained checkpoint.

  2. Marginal Heuristic Improvement: The algorithm relies on straightforward vector alignment and cosine similarity between a candidate's gradient descent direction and net parameter updates. Because gradient-to-update alignment is already a well-trodden concept in prior literature (e.g., Nasr et al., 2019; Zari et al., 2021), restricting this standard technique to MoE router rows represents an incremental heuristic rather than a strong, innovative methodological breakthrough.

  3. Writing and Presentation Can Be Improved: The manuscript suffers from dense notation, fragmented scoring modes (ROUTETRACE-S, ROUTETRACE-S-CAL, ZERO-BR), and convoluted descriptions that obscure core mechanics and hinder reproducibility. Specifically, readability is heavily compromised by broken mathematical equations (such as mismatched brackets and typos like in_{8} in Equation 5, and missing argument fields in Equation 7), stray quotation marks on page 3, and acronym bloat across multiple analysis states (B,H,PB, H, P).

Questions for authors
  1. Given that model developers rarely release both the raw base checkpoint and the post-trained checkpoint with aligned internal router structures in practice, what realistic threat model or enterprise scenario justifies assuming white-box paired checkpoint access?

  2. Since gradient-to-update alignment is a well-studied technique in white-box membership inference and federated learning, what fundamental theoretical insight does applying it to Mixture-of-Experts (MoE) router rows provide, and why is the scope restricted exclusively to MoEs?

  3. In Section 3.1, the gradient is computed at a hybrid state combining the base router with post-trained Q/VQ/V adapters. Can the authors justify why evaluating gradients at this intermediate state is mathematically necessary, and quantify how performance degrades if gradients are computed directly at the base or fully post-trained checkpoints?

  4. The proposed method relies on 3 shadow checkpoints, whereas baseline methods like LiRA and RMIA utilize 16 reference checkpoints. Because increasing auxiliary model counts heavily boosts attacker power, why was ROUTETRACE not evaluated under an equalized reference budget (e.g., using 16 shadows) to determine if its ranking advantage over output baselines persists at scale?

1 comment
AuthorsOpenReview ↗

Response on threat model and comparisons

Thank you for the questions.

Threat model. RouteTrace targets internal audits that retain base and post-trained checkpoints, and open-weight releases that publish both versions. It does not apply to a black-box API alone. We will state this scope more directly.

Relation to prior gradient alignment. Gradient–update alignment itself is not our contribution. RouteTrace uses the additional structure exposed by MoE routing: each token selects specific expert rows, allowing selected rows to be compared with equal-size boundary and random rows. Selected-row AUC exceeds these controls by 0.1715–0.3195. The new element is this routed-row comparison, not gradient alignment itself.

Hybrid state. Using post-trained Q/V representations raises mean router AUC from .8621 to .9413 on Qwen and from .8392 to .9618 on OLMoE. Replacing the base router with the post-trained router then adds only .0019 and .0058. Thus the main difference comes from the adapted representations, not the router evaluation point.

Reference budget. We did not train 16 RouteTrace shadows. With three total auxiliary checkpoints, RouteTrace has higher AUC than RMIA; with four, RMIA leads. At 16 references, both LiRA and RMIA have higher AUC than RouteTrace. We therefore present RouteTrace as a three-shadow operating point, not the highest-accuracy attack.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.