- Rating
- 2
- Confidence
- 1
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies label leakage in SL for LLM fine-tuning. The authors propose Gradient Mirage, which keeps that gradient useful for training but makes it inconsistent with the true labels in objective, scale and direction. It does this by combining dual-track backpropagation, an entropy-based masked loss, random per-token loss scaling, direction-only von Mises–Fisher noise and client-side gradient rescaling. Experimental results on Llama-2-7B, Llama-3-8B and DeepSeek-7B across CodeAlpaca, PIQA and GSM8K show much less label recovery.
Critical strengths
The paper identifies an important and underexplored privacy risk in LLM split learning (SL). The server can recover labels from activation gradients it already receives in normal training, and larger batches do not reduce this leakage. This gives the work clear practical motivation.
The method is simple and effective, and the results are promising. The client trains on the true loss but sends a gradient from a separate, randomized loss, which separates trainability from label privacy. On llama-3/CodeAlpaca, reconstruction ROUGE-L falls from 1.000 to 0.202 while perplexity barely changes. The proposed method beats every baseline in all nine settings, even when each defense is scored against its stronger reported attack. The evaluation spans three models, three datasets and three seeds, with ablations, adaptive attacks, sensitivity sweeps and task metrics
Critical weaknesses
This work only tested on SFT. Does it work for RL? nowadays senstive data are usually utilized at post training stage.
The evaluation is limited to three dense 7–8B models with very similar architectures (two Llama variants and one Deepseek), so model diversity is limited. It is unclear whether the privacy–utility trade-off holds for larger models, for popular MoE models, or for other architecture families. At least one larger model, ideally from a different family such as Qwen MoE (e.g., 30B-A3b), would make the claims much stronger.
Utility is measured almost entirely by ppl/rougel/meteor, which do not reliably reflect downstream performance. The only task-level results (HumanEval, GSM8K and PIQA in Fig. 15) are shown only as a figure, cover only the joint forward-and-backward protection setting, and do not compare against baselines. Task accuracy for Gradient Mirage and all baselines under the main setting (Table 1) is needed to support the claim that utility is preserved.
Questions for authors
How can this work transfer to RL post training?
Any evidence of industry adoption of SL?
The semi-white-box access session states that it is realistic and width adopted. any citation to support this claim?
7 comments
Response to Reviewer 5gwG (Part 2/3)
Task-Level Utility Evaluation (Acc %, ↑)
Llama3-8B
| Method | CodeAlpaca | PIQA | GSM8K |
|---|---|---|---|
| GP | 32.93±1.61 | 11.99±0.60 | 32.60±0.60 |
| GD | 38.82±2.88 | 62.80±1.26 | 18.45±0.73 |
| LDP(10240) | 34.76±1.61 | 59.50±5.74 | 18.47±0.46 |
| LDP(12288) | 36.59±2.20 | 63.28±6.04 | 18.80±0.40 |
| Top-Only | 38.82±5.67 | 56.11±0.52 | 16.93±0.43 |
| Standard | 43.90±2.20 | 59.01±3.94 | 48.52±0.27 |
| Ours(512) | 44.92±1.27 | 66.54±2.66 | 45.56±0.08 |
| Ours(1024) | 44.72±2.31 | 64.47±3.46 | 46.22±0.23 |
Llama2-7B
| Method | CodeAlpaca | PIQA | GSM8K |
|---|---|---|---|
| GP | 10.57±2.75 | 30.23±0.49 | 14.94±0.30 |
| GD | 14.02±1.61 | 31.34±0.59 | 2.60±0.52 |
| LDP(10240) | 11.79±0.70 | 47.39±0.49 | 2.50±0.80 |
| LDP(12288) | 12.80±0.61 | 49.35±0.63 | 3.11±1.25 |
| Top-Only | 13.41±2.20 | 30.05±0.41 | 2.98±0.16 |
| Standard | 16.87±0.93 | 46.30±1.15 | 19.18±0.35 |
| Ours(512) | 15.04±0.35 | 41.93±0.82 | 15.52±0.42 |
| Ours(1024) | 17.07±0.61 | 50.22±0.52 | 17.26±0.31 |
DeepSeek-LLM-7B
| Method | CodeAlpaca | PIQA | GSM8K |
|---|---|---|---|
| GP | 11.59±1.22 | 9.59±0.27 | 6.67±4.27 |
| GD | 20.53±1.96 | 41.78±1.12 | 3.44±0.79 |
| LDP(8192) | 22.36±1.96 | 47.66±2.01 | 6.09±0.24 |
| LDP(10240) | 23.37±1.53 | 48.93±1.11 | 6.14±0.13 |
| Top-Only | 20.33±2.31 | 32.72±1.03 | 4.88±0.74 |
| Standard | 26.22±0.61 | 41.51±0.63 | 20.22±0.43 |
| Ours(512) | 23.78±0.61 | 48.69±1.36 | 17.87±0.36 |
| Ours(1024) | 24.19±0.93 | 49.84±1.02 | 18.80±0.23 |
Response to Reviewer 5gwG (Part 3/3)
Q2: Industrial Adoption of Split Learning
We sincerely thank you for raising this important question. While large-scale production deployments of Split Learning (SL), particularly for LLM fine-tuning, are not yet extensively documented publicly, there is growing industrial interest and practical framework support. For example, NVIDIA has incorporated SL examples into its FLARE framework, demonstrating its applicability to collaborative training across distributed sites.
We acknowledge that the large-scale industrial adoption of Split Learning, particularly in LLM fine-tuning, is still at an early stage. However, even if a technology has not yet been widely deployed, proactively identifying and addressing its security vulnerabilities remains of great importance. We believe that our work contributes to this goal by investigating gradient-based privacy risks and developing effective defenses, thereby helping establish a more secure foundation for the future practical adoption of LLM-SL.
Furthermore, studying privacy issues in SL is essentially studying the privacy of hidden states under white-box or semi-white-box settings. This is the core problem in privacy-preserving inference with fully cloud-hosted models without relying on Trusted Execution Environments (TEEs). In our future work, we plan to shift our focus toward more practically relevant research problems. We sincerely thank you for pointing this out.
Q3: Justification for the Semi-White-Box Threat Model
We sincerely thank you for raising this question.
The semi-white-box setting has been widely adopted in existing studies on GMA and GMA-SL, including BiSR (Unveiling the Vulnerability of Private Fine-Tuning in Split-Based Frameworks for Large Language Models: A Bidirectionally Enhanced Attack, CCS'24), Gradient Dropout (Safeguarding Federated Learning From Data Reconstruction Attacks via Gradient Dropout, TIFS'26), LAMP (LAMP: Extracting Text from Gradients with Language Model Priors, NeurIPS'22), and the seminal work DLG (Deep Leakage from Gradients, NeurIPS'19).
One practical reason is that the server needs to load the Trunk model, meaning that it has access to the architecture and weights of the Trunk. By comparing them with publicly available open-source models, the server can identify the corresponding pretrained model and consequently obtain the initial weights of the Bottom and Top segments.
Furthermore, white-box assumptions are also adopted in privacy-preserving inference with fully cloud-hosted models, as seen in TWIST (Private Tenant-Dedicated LLM Serving via Joint Model-Token Transformation, CCS'26) and Understanding the Security Boundary of Obfuscation-based On-Device LLM Protection (CCS'26).
We apologize for not clearly explaining the semi-white-box assumption in our original manuscript. To clarify, the attacker does not have access to the dynamically updated parameters of the Top segment during training. Instead, the attacker only has access to a snapshot of its initial parameters, which serves as a proxy for conducting gradient-matching attacks.
Closing Remarks
We sincerely believe that you have carefully read our work and raised valuable suggestions and well-founded concerns. If there is anything that requires further clarification or additional experiments, please do not hesitate to let us know. We will do our very best to address them during the rebuttal period.
If our responses and additional results can adequately address your concerns, we would be deeply grateful if you could kindly reconsider your evaluation of our work and consider an upward revision of your score. Your recognition and support truly mean a great deal to us.
Response to Reviewer 5gwG (Part 1/3)
Dear Reviewer,
Thank you very much for your careful evaluation of our manuscript and for sharing your insightful comments. We greatly value your suggestions, which have helped us further improve the presentation and strengthen our work. We have carefully considered the points you raised and provide our detailed responses below.
Response
We fully agree that privacy protection during RL-based post-training deserves careful investigation in real-world applications.
In RL, the supervision signal usually does not come from a deterministic label sequence, but rather from reward rules or reward models. In this case, gradients are more likely to leak the reward model's score for a given sample rather than the data itself. Therefore, to the best of our knowledge, the research community has not yet established a unified evaluation standard for gradient leakage in RL. We have also mentioned the future research prospects of split-based RL in our Future Work section.
Regarding RL-based post-training in the split learning setting, we have some preliminary ideas that may be worth exploring. Our "Top-Only" training baseline could potentially be combined with a recent model obfuscation technique, TWIST: Private Tenant-Dedicated LLM Serving via Joint Model-Token Transformation (CCS'26), to provide a certain degree of privacy protection. Another version with a clearer description of the implementation is presented in the technical report Towards Privacy-Preserving LLM Inference via Covariant Obfuscation. Specifically, TWIST's model obfuscation technique could be used to obfuscate the entire Trunk to defend against privacy attacks targeting hidden states, while only the Top segment is fine-tuned during RL training. Accordingly, the number of Top layers would need to be increased, while the Bottom segment could contain only one layer.
Our work primarily focuses on the SFT setting, but we believe your suggestion is highly relevant to practical applications. We sincerely appreciate you for bringing up this important direction.
We sincerely apologize for giving you the impression that our experiments lack generality across different model architectures and parameter scales.
We have now supplemented our experiments with your suggested Qwen3-30B-A3B, as well as two additional models, Qwen2.5-14B and Qwen3-14B. The corresponding results have been added to Table 1. These experiments were conducted using a learning rate of 1e-4, a batch size of 8, gradient accumulation of 8, and the AdamW optimizer, on NVIDIA RTX PRO 6000 Blackwell GPU (96 GB). In addition, we have supplemented the defense results of these three models against LAMP and GMA-SL with cosine distance.
When we initially conducted this work, we did not have access to sufficient computational resources to perform these larger-scale experiments. We are now able to provide these additional results to address your concern, and we sincerely apologize for this limitation in our original submission. We have thoroughly updated Table 1 accordingly on page 8, and the experimental settings for the newly added models are detailed in Section 4.2. All newly added content is highlighted in red in the revised PDF for ease of reference. Given the size of the table, we sincerely appreciate your understanding in referring to the revised PDF for the full table.
We can observe that in the vast majority of cases, our method delivers both superior utility and stronger defense effectiveness compared to the baselines.
We sincerely apologize for this omission. We had initially assumed that perplexity (PPL) would be more indicative of performance, and that accuracy (Acc) might be somewhat unstable; we have now supplemented the Task-Level Utility Evaluation for both the original models and the newly added models. The corresponding results have been incorporated into Table 1. **We also present the task-level utility results here: