- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies membership inference risks arising from router telemetry exposed during inference in fine-tuned MoE models. The authors concatenate conventional output-side MIA features with aggregated routing features, and use same-domain auxiliary data to train shadow models and a membership classifier.
Across 3 MoE architectures × 3 data domains, adding router telemetry improves TPR@1% FPR by approximately 2.7--9.4 percentage points over the output-only ensemble. The authors further evaluate frozen-router training, LoRA, instruction tuning, discrete expert selection, partial telemetry, and reduced shadow-model access. The mechanistic analysis suggests that the leakage primarily comes from membership signals encoded in hidden representations after fine-tuning, while the router exposes a projection of these signals. The paper also evaluates simple defenses including noise, dropout, and quantization, and observes a privacy--utility trade-off.
Critical strengths
-
The paper shows that router telemetry should not be treated as inherently benign operational metadata, as it may create an additional membership leakage surface. This is a practically meaningful problem setting.
-
The experimental controls are reasonably comprehensive. The authors observe additional leakage across all 3×3 model--domain settings, and use frozen-router, LoRA, discrete-telemetry, and limited-shadow experiments to rule out several simple alternative explanations. In particular, the frozen-router result is convincing.
Critical weaknesses
-
The attack's practical realizability is not established. External API users receive neither full logits nor router telemetry. Operators who hold the telemetry can access the training data, so the attack adds nothing for them. The headline gains rely on full continuous expert loads, which no real service exposes.
-
Shadow and target models match in base checkpoint, domain, and fine-tuning recipe, while the threat model assumes no knowledge of the victim's optimizer, learning rate, training duration, data scale, or adaptation strategy. The results show that telemetry contains membership signal, they do not establish a practical attack.
-
The output-side ensemble reaches AUROC above 0.99 in eight of nine settings, so membership is nearly trivial from outputs alone.
-
TPR@1% FPR pools predictions from four held-out targets with no per-target results, confidence intervals, or seed repeats. The four targets share shadow models under leave-one-out, so the nine settings are not independent. The smallest gain, 2.7 points, is indistinguishable from noise without variance estimates.
-
The defense evaluation applies noise, dropout, and quantization and finds that leakage falls as fidelity falls. This outcome follows from the design, since the attack consumes the telemetry itself. Utility is measured as near-perfect correlation with the original loads, a bar stricter than any operational need, and alternatives such as batch aggregation or calibrated differential privacy are never compared.
-
Prior work already shows that routing signals expose input properties, and GateBleed performs membership inference through a routing side channel. The delta here is aggregating expert loads into an existing MIA ensemble, which is a feature-engineering contribution.
Questions for authors
-
Does the router-augmented MIA retain consistent gains when the shadow and victim models differ in learning rate, training duration, data scale, or adaptation strategy? This would directly affect my assessment of practical exploitability.
-
Could you report per-target TPR@1% FPR / AUROC together with bootstrap confidence intervals, to verify that the smaller gains are not driven by target-level variance hidden by pooling?
-
Do you have results on substantially larger MoE models, or at least an argument for why the current findings should be expected to generalize to much larger MoE deployments?
3 comments
Response to Reviewer XNyc (Part II)
W4 & Q2. Per-Target Results and Statistical Robustness
We conducted additional per-target analyses across all 36 held-out targets from the nine model and domain settings. For each target, we report TPR@1%FPR, AUROC, and paired bootstrap 95% confidence intervals for the improvements obtained by incorporating router telemetry, using 1,000 bootstrap replicates.
The TPR@1%FPR improvement is positive for all 36 targets, with 28/36 target-level confidence intervals excluding zero. AUROC improvements have confidence intervals excluding zero for 35/36 targets. When predictions are pooled within each model and domain setting, both improvements have confidence intervals excluding zero in all nine settings.
The following table presents the complete per-target results for OLMoE across all three data domains.
| Domain | Target | TPR@1%FPR (A_out → +Router) | ΔTPR [95% CI] | AUROC (A_out → +Router) | ΔAUROC [95% CI] |
|---|---|---|---|---|---|
| arXiv | M1 | 0.8696 → 0.9676 | +0.0980 [0.0689, 0.1301] | 0.9914 → 0.9979 | +0.0065 [0.0051, 0.0082] |
| M2 | 0.8764 → 0.9644 | +0.0880 [0.0653, 0.1095] | 0.9920 → 0.9980 | +0.0060 [0.0047, 0.0075] | |
| M3 | 0.8604 → 0.9344 | +0.0740 [0.0575, 0.0983] | 0.9887 → 0.9971 | +0.0083 [0.0069, 0.0099] | |
| M4 | 0.8680 → 0.9528 | +0.0848 [0.0594, 0.1042] | 0.9931 → 0.9980 | +0.0049 [0.0037, 0.0060] | |
| Pooled | 0.8685 → 0.9527 | +0.0842 [0.0745, 0.0995] | 0.9913 → 0.9974 | +0.0061 [0.0054, 0.0068] | |
| GitHub | M1 | 0.9233 → 0.9360 | +0.0127 [-0.0176, 0.0423] | 0.9951 → 0.9970 | +0.0019 [0.0007, 0.0033] |
| M2 | 0.8420 → 0.9027 | +0.0607 [0.0020, 0.1563] | 0.9919 → 0.9955 | +0.0036 [0.0019, 0.0058] | |
| M3 | 0.7840 → 0.9153 | +0.1313 [0.0345, 0.2253] | 0.9928 → 0.9960 | +0.0033 [0.0018, 0.0050] | |
| M4 | 0.8520 → 0.9000 | +0.0480 [-0.0150, 0.0995] | 0.9940 → 0.9958 | +0.0019 [0.0004, 0.0035] | |
| Pooled | 0.8585 → 0.9133 | +0.0548 [0.0201, 0.0843] | 0.9933 → 0.9960 | +0.0027 [0.0018, 0.0036] | |
| Wikipedia | M1 | 0.9420 → 0.9653 | +0.0233 [0.0053, 0.0511] | 0.9981 → 0.9987 | +0.0007 [0.0002, 0.0011] |
| M2 | 0.9353 → 0.9533 | +0.0180 [-0.0066, 0.0447] | 0.9974 → 0.9984 | +0.0010 [0.0005, 0.0016] | |
| M3 | 0.9373 → 0.9573 | +0.0200 [-0.0096, 0.0485] | 0.9976 → 0.9982 | +0.0006 [0.0000, 0.0011] | |
| M4 | 0.9407 → 0.9813 | +0.0407 [0.0121, 0.0744] | 0.9977 → 0.9992 | +0.0015 [0.0006, 0.0026] | |
| Pooled | 0.9373 → 0.9655 | +0.0282 [0.0146, 0.0399] | 0.9977 → 0.9986 | +0.0009 [0.0006, 0.0012] |
Across all three domains, all 12 OLMoE targets exhibit positive TPR@1%FPR improvements, with eight target-level confidence intervals excluding zero. Notably, the pooled improvements remain positive with confidence intervals above zero in all three domains.
The wider confidence intervals for some GitHub and Wikipedia targets reflect the limited number of non-member examples at the 1% FPR operating point. Each target contains 1,500 non-members, corresponding to approximately 15 false positives at this threshold. The paired bootstrap results quantify evaluation-sample uncertainty conditional on the trained models, while the per-target results reveal variation across independently fine-tuned targets.
Together, these analyses provide additional evidence that the observed improvements are consistent across held-out targets and are not solely driven by a small number of favorable target models.
W5. Interpretation of the Telemetry Defense Experiments
Section 5.6 quantitatively evaluates the privacy and telemetry-fidelity trade-off under Gaussian noise, dropout, and quantization. We use two complementary fidelity metrics: Pearson correlation measures preservation of expert loads, while top-1 expert agreement measures preservation of the most heavily loaded expert (Appendix F).
For Gaussian noise at , the additional TPR@1%FPR gain decreases from 0.0841 to 0.0159, while correlation and top-1 agreement drop to 0.588 and 0.407. At dropout rate , the gain falls to 0.0040, with correlation of 0.047 and top-1 agreement of 0.050. Quantization at yields a gain of 0.0079, with correlation of 0.045 and top-1 agreement of 0.076.
These results quantify the substantial loss of routing fidelity associated with suppressing membership leakage across three perturbation mechanisms. They provide concrete privacy and fidelity operating points, while task-specific monitoring utility and alternative mechanisms such as batch aggregation and differential privacy remain valuable directions for further evaluation.
Response to Reviewer XNyc (Part III)
W6. Distinction from Prior Routing Privacy Work
Our contribution addresses a distinct privacy question: whether routing information reveals membership in the fine-tuning dataset of a deployed MoE model. Earlier routing-related privacy work primarily investigates inference-time input properties, including prompt information and expert choices. GateBleed demonstrates a hardware timing side channel, including membership inference in a separate model setting and MoE expert-choice recovery.
Our study directly evaluates router telemetry as a fine-tuning membership signal, measures its incremental value against a strong ensemble of output-based MIA features, and characterizes how the signal changes across fine-tuning regimes, telemetry restrictions, and shadow-model configurations. The frozen-router and matched-rank projection experiments further identify a mechanism through which updated hidden representations can expose membership-related information through routing. Together, these findings establish a distinct privacy risk and provide a systematic empirical and mechanistic account of its behavior.
Q3. Generalization to Larger MoE Models
Our evaluation covers three MoE architectures with different expert configurations and routing structures, consistently demonstrating additional membership leakage from router telemetry. The frozen-router analysis further reveals a general mechanism: fine-tuning changes hidden representations, allowing even unchanged router parameters to expose membership-related information through expert routing decisions.
This mechanism is rooted in the fundamental routing operation shared by sparse MoE architectures, including larger models that use learned expert selection. The consistency across architectures and fine-tuning regimes provides evidence that the observed leakage is associated with general MoE routing behavior. While its magnitude may vary with model scale and configuration, our findings establish a mechanistic basis for understanding this privacy risk beyond the specific models evaluated.
Response to Reviewer XNyc (Part I)
We thank the reviewer for recognizing the importance of router telemetry as a privacy surface and the breadth of our experimental controls. We provide further evidence addressing the deployment assumptions, transferability, statistical robustness, and scope of our findings.
W1. Practical Realizability and Telemetry Access
Our threat model considers an adversary with access to request-associated inference telemetry and token-level prediction scores. This includes observers with scoped monitoring or tracing permissions, infrastructure insiders, and actors who obtain access to internal observability interfaces. Serving logs, model parameters, and private fine-tuning datasets can be protected by separate access controls. Accordingly, access to routing traces can create a meaningful membership inference risk even when the underlying training records remain inaccessible.
Router information is available at concrete internal execution interfaces. Hugging Face Transformers supports
output_router_logits=Truefor supported MoE models (https://huggingface.co/docs/transformers/v4.47.1/model_doc/qwen2_moe). NVIDIA TensorRT provides MoE tensors such asselectedExpertsForTokensandscoresForSelectedExperts(https://docs.nvidia.com/deeplearning/tensorrt/latest/inference-library/transformers-moe.html). Our evaluated setting requires traces linked to individual requests and access to token-level prediction scores through a scoring or instrumentation interface. We will specify these permissions explicitly in the threat model.Importantly, the additional leakage persists with discrete expert selections. Table 3 reports positive TPR@1%FPR gains in all nine model and domain settings using expert-selection frequencies, ranging from +0.0283 to +0.1043. This substantially reduces the precision required of exposed routing information. The results establish a concrete privacy risk for systems that expose the specified observations.
W2 & Q1. Robustness to Shadow and Target Training Mismatch
We performed additional experiments on OLMoE with arXiv using 49 shadow and target configuration pairs. The seven configurations cover full fine-tuning, frozen-router training, LoRA, and five learning rates between 5e-6 and 1.2e-5. Both attack variants are trained on the shadow configuration and evaluated on held-out targets. Adding routing telemetry improves TPR@1%FPR in all 49 pairs.
The following results fix the target to the original full fine-tuning setting (LR=8e-6) and vary the shadow training configuration.
The positive gains under LoRA, frozen-router training, and learning-rate mismatch demonstrate that the attack retains its added value across substantially different shadow training configurations. The present matrix controls the architecture, data domain, and one-epoch training duration, allowing the effects of adaptation choice and learning rate to be evaluated directly.
W3. Additional Leakage in More Challenging MIA Settings
Our new experiments also cover target configurations where output-based inference is considerably more difficult. Using shadow models trained with the original LR=8e-6, we evaluated full fine-tuning targets trained at lower learning rates.
At LR=5e-6, adding router telemetry improves TPR@1%FPR by 11.7 percentage points even though output-only AUROC is 0.905. At LR=6e-6, the gain is 13.4 percentage points. These results extend the main empirical finding to settings with substantially greater headroom for membership inference.