acceptodds
Under review as a conference paper at ICLR 2027

Gradient Mirage: Trainable yet Label-Unidentifiable Gradients in Large Language Model Split Learning

Abstract

Gradient matching attacks (GMAs) in LLM split learning (SL) rely on a critical yet underexplored assumption: the gradient exposed at the split interface is a faithful derivative of the client’s full-label training objective. This gradient-label correspondence allows a curious server to recover private labels by searching for a sequence whose induced gradient explains the observation. We propose Gradient Mirage, a defense that breaks this correspondence without discarding the optimization utility of the backward signal. Our key idea is to induce the adversary to solve a misspecified inverse problem, in which no plausible label sequence in the sequence space can explain the observed gradients. Concretely, Gradient Mirage achieves this by inducing inconsistency across three dimensions: objective, direction, and scale. Selective Autoregressive Supervision derives the exposed gradient from a masked surrogate loss rather than the full-label objective assumed by the attacker; Scale Blinding then applies randomized multiplicative rescaling, obscuring the gradient's natural magnitude; and Directional Privatization further randomizes the gradient direction while preserving its magnitude through the von Mises–Fisher (vMF) mechanism under a directional metric differential privacy guarantee. Crucially, utility is preserved: the Top segment still learns from all target tokens via Dual-Track Backpropagation, the exposed gradient remains informative since each supervised token retains its complete autoregressive context, and Bottom-Gradient Recovery restores the effective gradient for Bottom-segment optimization. Extensive experiments show that Gradient Mirage provides substantially stronger protection than existing defenses under comparable fine-tuning performance, achieving a better privacy-utility trade-off.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Split Learning, Gradient Leakage, Gradient Matching Attack, Privacy Preserving

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_5gwGOpenReview ↗
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
AuthorsOpenReview ↗

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

W1/Q1: Generalization to RL Post-Training

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.

W2: Limited Model Diversity and Scalability to Larger/MoE Models

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.

W3: Insufficient Downstream Task-Level Utility Evaluation

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:

AuthorsOpenReview ↗

Response to Reviewer 5gwG (Part 2/3)

Task-Level Utility Evaluation (Acc %, ↑)

Llama3-8B

MethodCodeAlpacaPIQAGSM8K
GP32.93±1.6111.99±0.6032.60±0.60
GD38.82±2.8862.80±1.2618.45±0.73
LDP(10240)34.76±1.6159.50±5.7418.47±0.46
LDP(12288)36.59±2.2063.28±6.0418.80±0.40
Top-Only38.82±5.6756.11±0.5216.93±0.43
Standard43.90±2.2059.01±3.9448.52±0.27
Ours(512)44.92±1.2766.54±2.6645.56±0.08
Ours(1024)44.72±2.3164.47±3.4646.22±0.23

Llama2-7B

MethodCodeAlpacaPIQAGSM8K
GP10.57±2.7530.23±0.4914.94±0.30
GD14.02±1.6131.34±0.592.60±0.52
LDP(10240)11.79±0.7047.39±0.492.50±0.80
LDP(12288)12.80±0.6149.35±0.633.11±1.25
Top-Only13.41±2.2030.05±0.412.98±0.16
Standard16.87±0.9346.30±1.1519.18±0.35
Ours(512)15.04±0.3541.93±0.8215.52±0.42
Ours(1024)17.07±0.6150.22±0.5217.26±0.31

DeepSeek-LLM-7B

MethodCodeAlpacaPIQAGSM8K
GP11.59±1.229.59±0.276.67±4.27
GD20.53±1.9641.78±1.123.44±0.79
LDP(8192)22.36±1.9647.66±2.016.09±0.24
LDP(10240)23.37±1.5348.93±1.116.14±0.13
Top-Only20.33±2.3132.72±1.034.88±0.74
Standard26.22±0.6141.51±0.6320.22±0.43
Ours(512)23.78±0.6148.69±1.3617.87±0.36
Ours(1024)24.19±0.9349.84±1.0218.80±0.23
AuthorsOpenReview ↗

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.

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

This paper proposes Gradient Mirage to protect private labels against gradient matching attacks in LLM split learning by manipulating the gradients exposed to the server. It introduces selective autoregressive supervision (SAS), scale blinding, and directional privatization to reduce privacy leakage, with dual-track backpropagation and bottom-gradient recovery to retain model utility. The experimental results across multiple datasets and models demonstrate a better privacy-utility trade-off than the baseline methods.

Critical strengths
  1. Effective privacy protection with retained model utility: In the experiments, the proposed method reduces BiSR(b)'s reconstruction performance while retaining model utility, demonstrating a better privacy-utility trade-off than the baseline approaches.
  2. Extensive experiments and ablation studies: The experiments cover multiple datasets and models, with ablation studies showing the defense's effectiveness and the claimed contributions of each component.
Critical weaknesses
  1. Limited attack strategy diversity: The evaluation mainly uses BiSR(b) as the attack approach. The experimental results provide evidence within this attack framework, but it remains unclear whether the proposed defense generalizes to different attack strategies.
Questions for authors
  1. Complementarity of defense components: The ablation studies show benefits from individual components that reduce reconstruction quality or retain model utility. Do these mechanisms provide complementary benefits and enhance the overall privacy-utility trade-off when combined, or do their effects overlap?
  2. Hyperparameter sensitivity and selection: The defense involves several hyperparameters across different components. How sensitive are defense performance and model utility to the choice of these hyperparameters? Is there practical guidance for selecting them?
  3. Model-dependent utility impact: Table 1 shows near-preserved PPL on LLaMA-3 but larger degradation on the LLaMA-2 and DeepSeek models. Is there an explanation or intuition behind this difference?
5 comments
AuthorsOpenReview ↗

Response to Reviewer NLAM (Part 1/3)

Dear Reviewer,

We sincerely thank you for carefully reading our paper and giving us a positive rating. We appreciate your constructive suggestions and insightful questions. We will do our best to address each of your concerns in detail below.

Response

W1: Limited Attack Strategy Diversity

We would like to clarify that BiSR(b) is essentially an attack paradigm rather than a single fixed attack strategy. Its core lies in the choice of the gradient matching objective. In the original BiSR paper, the best-performing matching objective was the weighted combination of L1 and L2 distances adopted from TAG.

Beyond this setting, we have also included GMA-SL using cosine distance as the matching objective in our appendix. In addition, there is another GMA-SL attack adapted from LAMP by the original authors of BiSR, which they directly refer to as LAMP. However, the original LAMP was designed for gradient attacks against LLM fine-tuning in federated learning. We have also reproduced the implementation from their official repository and added the corresponding defense results against LAMP in Appendix D.5. We present a simplified version of these results here, and the complete table can be found in the revised PDF.

Reconstruction under LAMP (mean ± std over three seeds; lower is better).

CodeAlpaca

ModelMethodROUGE-L F1 ↓ROUGE-2 F1 ↓METEOR ↓
Llama3-8BOurs(512).04±.01.00±.00.03±.00
Llama3-8BOurs(1024).09±.02.01±.01.06±.02
Llama2-7BOurs(512).08±.01.01±.00.06±.01
Llama2-7BOurs(1024).12±.01.01±.00.09±.01
DeepSeek-LLM-7BOurs(512).08±.02.02±.01.09±.01
DeepSeek-LLM-7BOurs(1024).06±.01.04±.01.08±.01
Qwen2.5-14BOurs(512).19±.03.08±.02.17±.02
Qwen2.5-14BOurs(1024).24±.06.11±.07.21±.05
Qwen3-14B-BaseOurs(512).17±.01.05±.01.14±.02
Qwen3-14B-BaseOurs(1024).19±.02.09±.01.21±.01
Qwen3-30B-A3B-BaseOurs(512).12±.05.01±.00.06±.01
Qwen3-30B-A3B-BaseOurs(1024).17±.03.01±.00.09±.01

PIQA

ModelMethodROUGE-L F1 ↓ROUGE-2 F1 ↓METEOR ↓
Llama3-8BOurs(512).03±.01.00±.00.02±.00
Llama3-8BOurs(1024).06±.01.00±.00.05±.01
Llama2-7BOurs(512).04±.00.00±.00.04±.00
Llama2-7BOurs(1024).10±.01.01±.00.09±.01
DeepSeek-LLM-7BOurs(512).09±.01.00±.00.07±.01
DeepSeek-LLM-7BOurs(1024).06±.02.01±.00.12±.00
Qwen2.5-14BOurs(512).08±.01.01±.01.06±.01
Qwen2.5-14BOurs(1024).17±.02.02±.01.12±.02
Qwen3-14B-BaseOurs(512).10±.01.01±.01.07±.01
Qwen3-14B-BaseOurs(1024).22±.01.02±.01.14±.01
Qwen3-30B-A3B-BaseOurs(512).08±.01.01±.00.06±.01
Qwen3-30B-A3B-BaseOurs(1024).09±.01.01±.00.08±.02

GSM8K

ModelMethodROUGE-L F1 ↓ROUGE-2 F1 ↓METEOR ↓
Llama3-8BOurs(512).05±.01.00±.00.04±.01
Llama3-8BOurs(1024).10±.01.01±.00.06±.01
Llama2-7BOurs(512).08±.01.01±.01.06±.01
Llama2-7BOurs(1024).13±.01.02±.01.09±.01
DeepSeek-LLM-7BOurs(512).10±.01.02±.01.10±.01
DeepSeek-LLM-7BOurs(1024).09±.00.05±.01.11±.01
Qwen2.5-14BOurs(512).25±.00.12±.01.23±.01
Qwen2.5-14BOurs(1024).26±.03.12±.02.24±.02
Qwen3-14B-BaseOurs(512).26±.03.10±.02.23±.02
Qwen3-14B-BaseOurs(1024).23±.05.08±.03.19±.05
Qwen3-30B-A3B-BaseOurs(512).26±.01.11±.03.22±.01
Qwen3-30B-A3B-BaseOurs(1024).23±.01.08±.02.18±.03
AuthorsOpenReview ↗

Response to Reviewer NLAM (Part 2/3)

Furthermore, we have supplemented our evaluation with a stronger token-wise adaptive attack, in which an independently optimizable weight is introduced for each token during gradient matching. This allows the attacker to jointly optimize the reconstructed label sequence and token-wise scaling factors, making it more sophisticated than the global-scale adaptive attack considered in our original appendix. The corresponding results have been added to Table 2. We also show the results of this stronger attack here:

ModelεToken-wise ROUGE-L F1 ↓BiSR(b) ROUGE-L F1 ↓Δ ROUGE-L F1 (pp)Token-wise TRR (%) ↓BiSR(b) TRR (%) ↓Δ TRR (pp)
Llama3-8B128.05 ± .02.04+1.147.65 ± 2.085.00+2.65
256.10 ± .02.06+3.529.27 ± 2.127.00+2.27
384.13 ± .03.09+3.5012.17 ± 2.149.00+3.17
512.23 ± .03.20+2.8427.77 ± 2.2425.00+2.77
1024.36 ± .03.35+1.0239.91 ± 2.5537.00+2.91
Qwen2.5-14B128.15 ± .04.13+1.996.39 ± 7.026.00+0.39
256.19 ± .05.15+4.3610.44 ± 7.968.00+2.44
384.26 ± .05.18+7.6117.08 ± 7.7611.00+6.08
512.30 ± .03.24+5.6924.31 ± 5.1122.00+2.31
1024.34 ± .03.29+5.0933.03 ± 4.7826.00+7.03

We have also extended these evaluations to three additional models: Qwen2.5-14B, Qwen3-14B-Base, and Qwen3-30B-A3B-Base, providing further evidence of the generalizability of our defense across different attack strategies, model scales, and architectures, including the MoE architecture of Qwen3-30B-A3B-Base. The corresponding results have been added to Table 1.

Q1: Complementarity of Defense Components

We mainly demonstrate the complementarity of our defense components through the ablation experiments in Tables 3, 5, and 6.

As shown in Table 5, under cosine-distance-based GMA-SL, Scale Blinding alone achieves a ROUGE-L F1 of 0.55. After introducing SAS (ρ = 0.6), it decreases to 0.37, and further incorporating Directional Privatization (ε = 1024) reduces it to 0.17. Furthermore, as shown in Table 6, under the magnitude-sensitive BiSR(b) attack, removing Scale Blinding increases ROUGE-L F1 from 0.35 to 0.71. These experimental results demonstrate that the three components provide complementary privacy protection by introducing inconsistencies in the objective, direction, and scale of the exposed gradients.

More importantly, these three components can also reinforce each other against adaptive attacks targeting individual defense mechanisms. When only one component is applied, an attacker may adapt the attack strategy specifically to that mechanism, potentially bypassing its protection. However, when the three components are combined, even if an adaptive attack successfully compensates for one component, the other two components can still interfere with gradient matching from different aspects. Therefore, their combination may provide stronger resistance to adaptive attacks than any individual component alone.

Regarding training utility, Table 3 shows that removing either Dual-Track Backpropagation or Bottom-Gradient Recovery consistently increases PPL across all three datasets. This indicates that these two components help preserve training utility while the other components provide privacy protection.

Therefore, our ablation results provide empirical evidence that the proposed components complement each other in improving the overall privacy-utility trade-off, although some overlap between their effects may exist.

AuthorsOpenReview ↗

Response to Reviewer NLAM (Part 3/3)

Q2: Hyperparameter Sensitivity and Selection

We sincerely thank you for raising this important question. We have conducted additional experiments to investigate the sensitivity of our method to the privacy budget ε and provide practical guidance on hyperparameter selection.

Privacy Budget ε. We observed strong defense effectiveness across all evaluated models at ε = 512. To further investigate the privacy–utility trade-off, we evaluated smaller privacy budgets (ε = 128, 192, 256, and 384) on Llama3-8B and Qwen2.5-14B. The results are presented below:

ModelMetric128192256384
Llama3-8BPPL/Acc (%)2.546/43.292.523/38.412.513/35.372.498/44.51
ROUGE-L F ↓0.040.040.060.09
Qwen2.5-14BPPL/Acc (%)2.372/68.292.366/65.242.357/65.852.349/65.24
ROUGE-L F ↓0.130.140.150.18

The results show that smaller privacy budgets consistently reduce reconstruction effectiveness while maintaining relatively stable utility. Based on these observations, we recommend ε = 512 as a practical default that provides effective empirical protection across our evaluated models, while smaller values can be selected when stronger privacy protection is desired.

Scale-Blinding Coefficient m. We have also provided sensitivity analyses and recommended settings for the scale-blinding coefficient m in the Appendix. Furthermore, for larger values of m, our proposed Trunk-Frozen Training strategy can effectively preserve training utility, mitigating the potential performance degradation caused by stronger scale blinding.

Token Sampling Ratio ρ. For the token sampling ratio ρ in SAS, we recommend a practical heuristic based on the number of training epochs. Specifically, for multi-epoch training, ρ can be set approximately 20 percentage points above the reciprocal of the number of epochs, i.e., ρ ≈ 1/E + 0.2, where E denotes the total number of training epochs. For example, when training for 3 epochs, we recommend setting ρ between 0.5 and 0.6. This heuristic aims to maintain sufficient supervision coverage across training epochs while preserving randomness in token selection.

Overall, these results and practical recommendations provide guidance for selecting the key hyperparameters of our method to balance privacy protection and model utility.

Q3: Model-Dependent Utility Impact

Different models naturally have different architectures, pretrained weights, and optimization characteristics, so it is expected that they may respond differently to the same defense mechanism. Moreover, our method is fundamentally based on introducing perturbations into the training process to protect privacy, rather than providing completely lossless privacy protection. Therefore, some degree of utility degradation is unavoidable, and its magnitude may vary across different models.

As shown in our experiments, Gradient Mirage achieves different levels of utility preservation across Llama3-8B, Llama2-7B, and DeepSeek-LLM-7B. Nevertheless, our goal is not to guarantee identical or completely lossless utility across all models, but rather to achieve a reasonable privacy-utility trade-off while substantially reducing gradient leakage.

Furthermore, we have supplemented experiments on Qwen2.5-14B, Qwen3-14B-Base, and Qwen3-30B-A3B-Base to provide a more comprehensive assessment across different model architectures and scales. We have also supplemented our evaluation with downstream task-level utility results for all evaluated models. We kindly invite you to refer to these additional results for a more comprehensive evaluation of the practical utility of our method.

We have comprehensively revised Table 1 accordingly and further clarified the experimental settings for the newly added models in Section 4.2. All new additions are marked in red in the revised PDF for ease of reference.

Closing Remarks

We truly appreciate your careful review of our work and your insightful suggestions and constructive concerns. If any further clarification or additional experiments are needed, please do not hesitate to let us know. We will make every effort to address them during the rebuttal period.

We hope our responses and additional results have addressed your concerns. If so, we would be very grateful if you could consider increasing the overall rating and confidence score. Your support means a great deal to us. Once again, we sincerely thank you for your time, careful review, and valuable suggestions.

Reviewer_Ncv4OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

In this paper they propose defence, Gradient Mirage, for Gradient Matching Attacks (GMA). This paper tries to address the gradient-label correspondance by introducing objective, directional, and scale inconsistencies. In current form the paper is very unclear and difficult to follow.

Critical strengths
  • The goal of decoupling model learning from gradient leakage by introducing inconsistencies during training is interesting and seems novels.
  • The paper provides both theoretical and empirical analysis, although there is lack of clarity and connection for theory provided.
Critical weaknesses
  • Please explain the threat model and preliminary clearly. For example, there is not mention of top, bottom, and trunk parts here. Instead new terms are introduced here.
  • Furthermore, provide motivation for gradient mirage defence, along with justification for sub-components and diff optimisation tricks inside in this defence.
  • Very limited experimental setting details provided. With current details, reproducibility seems limited.
Questions for authors
  • The baselines (used in Table 1) is not explained and no note for lack of results for some cells in the Table 1.
  • I encourage authors to deeply think about the narrative and writing of the paper. Articulate the motivation of mirage defence clearly with all arguments explained well. Also, avoid jumping arguments and ensure consistency across.

Some minor comments:

  • Lot of unnecessary capitalisation across the paper.
  • Many of the figure and tables in the paper are too small.
  • A very difficult read. For ex: try to explain what is top, trunk, and bottom segment. The same could have been done when the threat model was defined.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Ncv4 (Part 1/2)

Dear Reviewer,

We sincerely appreciate your time and thoughtful feedback on our work. We have carefully addressed your concerns and revised the manuscript accordingly. Please find our point-by-point responses below.

Response

W1: Threat Model and Preliminaries

We originally explained these concepts in the second paragraph of Section 2.1. However, we acknowledge that our explanation may not have been sufficiently clear or intuitive. To address this, we have added Figure 2 and further textual clarifications in Sections 2.1 and 2.4, with all revisions highlighted in red. We sincerely apologize for the lack of clarity in the original presentation. We hope these revisions address your concerns and make our presentation clearer.

W2: Motivation for Gradient Mirage / Justification of Defense Components

We sincerely thank you for this constructive suggestion. We apologize that the motivation and design rationale were not sufficiently clear in our original presentation. Motivation. Existing forward-pass defenses, such as activation perturbation in DP-Forward (DP-Forward: Fine-tuning and Inference on Language Models with Differential Privacy in Forward Pass) and embedding obfuscation in ObfusLM (ObfusLM: Privacy-preserving Language Model Service against Embedding Inversion Attacks), can effectively defend against forward-pass attacks. However, they do not necessarily prevent label leakage through backward gradients. Even when activations are perturbed, the corresponding gradients are computed directly with respect to these perturbed activations and still encode label-dependent information. Consequently, an adversarial server can perform gradient-matching attacks using the same perturbed activations, potentially recovering private labels despite the injected noise. Meanwhile, directly perturbing gradients may significantly impair training utility. This motivates Gradient Mirage, which aims to disrupt gradient–label correspondence while preserving effective fine-tuning.

Defense Components. Each component addresses a distinct aspect of gradient matching: (i) Selective Autoregressive Supervision (SAS) introduces objective inconsistency by replacing full-label supervision with a masked surrogate loss for the exposed gradient. Entropy-aware grouping maintains supervision across different token difficulty levels. (ii) Scale Blinding randomizes token-wise gradient contributions to hinder magnitude-sensitive matching. (iii) Directional Privatization perturbs gradient directions using the vMF mechanism, further disrupting directional matching while preserving gradient norms. These components provide complementary protection across objective, scale, and direction.

Optimization Strategies. To mitigate the potential utility degradation, Dual-Track Backpropagation preserves full-label learning for the client-side Top segment, while Bottom-Gradient Recovery compensates for the average scale amplification introduced by Scale Blinding. The protected gradients also remain useful for Trunk optimization.

Experimental Justification. Our ablation studies support these design choices. Table 5 evaluates the effects of SAS and Directional Privatization; Table 6 shows that removing Scale Blinding increases ROUGE-L F1 from 0.35 to 0.71; and Table 3 confirms that removing either optimization strategy consistently increases fine-tuning perplexity. We greatly appreciate your feedback and have strengthened the design motivation and component-level explanations in Section 3. We hope these clarifications and supporting results adequately address your concerns.

AuthorsOpenReview ↗

Response to Reviewer Ncv4 (Part 2/2)

W3: Experimental Settings and Reproducibility

We apologize that the experimental configurations were not presented in sufficient detail, and we appreciate the opportunity to clarify them. Experimental Settings. As described in Section 4, our experiments cover three LLMs (Llama-2-7B, Llama-3-8B, and DeepSeek-LLM-7B) and three datasets (CodeAlpaca, GSM8K, and PIQA). All experiments use bfloat16 precision, with 6 decoder layers assigned to the Bottom segment, 5 to the Top segment, and the remaining layers to the Trunk. We use a batch size of 8 and repeat experiments with 3 random seeds {42, 43, 44}. For attack evaluation, we perform 600 fine-tuning steps, launching a GMA-SL attack every 50 steps using BiSR(b) with a gradient-matching weight of β = 0.65. For Gradient Mirage, we set the SAS sampling ratio to ρ = 0.6, use k = 3 token groups, sample Scale Blinding coefficients from Uniform[1500, 2000], and evaluate directional privacy budgets ε ∈ {512, 1024}. We report ROUGE-L F1 and METEOR for reconstruction quality and perplexity (PPL) for fine-tuning utility. Additionally, we have incorporated new evaluation metrics into Table 1 to provide a more comprehensive assessment of our method.

Reproducibility. We fully acknowledge your concern regarding reproducibility. To facilitate independent verification and reproduction of our results, we have uploaded our source code.

Q: Missing Results in Table 1

We apologize that the missing entries in Table 1 were not sufficiently explained.

Missing Results for Top-Only. The missing reconstruction results in Table 1 correspond to the Top-Only baseline, where only the client-side Top segment is fine-tuned, while the Bottom and Trunk segments remain frozen. Since no backward gradients are transmitted to the Trunk, the server cannot perform the considered gradient leakage attack. Therefore, reconstruction metrics are not applicable and are marked as “–”. However, restricting fine-tuning to the Top segment also limits model adaptation, resulting in degraded fine-tuning performance compared with full split learning, as reflected by the higher PPL values in Table 1. We have added this clarification in Section 4.2 (Benchmarking Gradient Mirage), immediately after the reference to Table 1, with the revised text highlighted in red. We sincerely appreciate your helpful feedback and hope these revisions address your concerns.

Closing Remarks

We sincerely thank you for your time and feedback. We have carefully considered your concerns and made corresponding revisions to improve the clarity, completeness, and reproducibility of our work. However, we respectfully feel that a score of 1 is disproportionately low and amounts to an almost complete dismissal of our work. While we acknowledge that the original presentation could be improved, we sincerely hope that these presentation issues will not overshadow the technical contributions and experimental evidence of our work. We respectfully ask you to carefully consider the significance of our contributions, the supporting experimental results, and our efforts to address the raised concerns. We would be deeply grateful if you could reconsider whether a score of 1 fairly reflects the merits of our work, particularly in light of the clarifications and revisions provided. Thank you again for your time and consideration.

I have gone through the rebuttal. Thanks for changes and additional experiments, most of my concerns are addressed. Some clarification and connection of theory with overall work and empirical results. I have increased by score for presentation, soundness and overall rating. I am still not satisfied with presentation of the paper and is not in the shape to be accepted. I see novelty of the work as supported by other reviewers.

Good luck.

AuthorsOpenReview ↗

Thank You for Your Thoughtful Feedback and Reconsideration

Dear Reviewer,

Thank you for taking the time to read our rebuttal and reconsider your scores. We're really glad that our additional experiments and explanations helped address your concerns. We truly appreciate your support and the opportunity to improve our work through your feedback.

Best regards,
The Authors

Reviewer_1zVLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Gradient Mirage, a gradient privacy protection method for large language model split learning, aiming to prevent servers from recovering clients' private label sequences through gradient matching attacks (GMAs). The authors identify the correspondence between shared gradients and the true label supervision objective as a key vulnerability exploited by such attacks. To disrupt this correspondence, Gradient Mirage introduces inconsistencies across three dimensions: objective, direction, and scale, through Selective Autoregressive Supervision (SAS), directional randomization, and randomized gradient scaling, respectively. Meanwhile, Dual-Track Backpropagation and Bottom-Gradient Recovery are employed to preserve model training performance. Experiments across multiple models and datasets demonstrate that the proposed method significantly reduces the quality of text reconstruction by gradient inversion attacks compared with existing defenses, while maintaining competitive training utility in most settings, achieving a favorable privacy–utility trade-off.

Critical strengths
  1. Clear and Meaningful Research Motivation with a Well-Structured Method Design. This paper addresses the problem of gradient-based label leakage in large language model split learning, identifying the correspondence between shared gradients and the true label supervision objective as a key vulnerability exploited by gradient matching attacks. Building on this insight, the authors propose a defense strategy that introduces inconsistencies across three dimensions: objective, direction, and scale. Dual-Track Backpropagation and Bottom-Gradient Recovery are further employed to balance privacy protection and model training utility. Overall, the research motivation is clear, and the proposed method is systematic and well aligned with the identified problem.
  2. Extensive Experimental Validation and Significant Privacy Protection Improvements. The authors conduct extensive experiments across multiple large language models and datasets, comparing the proposed method with existing gradient privacy protection approaches. The results demonstrate that Gradient Mirage significantly reduces text reconstruction quality under gradient inversion attacks while maintaining competitive training performance in most settings. Furthermore, ablation studies and parameter sensitivity analyses provide additional evidence of the effectiveness of the proposed components, supporting the claimed privacy–utility trade-off.
Critical weaknesses
  1. Insufficient Evaluation of Robustness Against Adaptive Attacks. The paper primarily evaluates its defense against existing gradient matching attacks. However, the proposed method fundamentally relies on disrupting the correspondence between gradients and labels assumed by the attacker. An adversary aware of the defense mechanism could potentially adapt the gradient matching strategy by jointly estimating supervision masks, token-wise random scaling factors, and other perturbations. Although the authors consider an adaptive attack in the appendix, it mainly focuses on optimizing a global scaling parameter. The effectiveness of the proposed method against more sophisticated adaptive attacks or multiple gradient observations remains insufficiently validated, limiting confidence in its robustness.
  2. Limited Connection Between Theoretical Privacy Guarantees and Practical Protection. The paper provides a Metric Differential Privacy guarantee over gradient directions through the vMF mechanism. However, this guarantee does not directly translate into privacy protection for the original text or label sequences. Furthermore, the proposed lower bound on the Amplitude Signal-to-Noise Ratio (ASNR) primarily reflects a geometric constraint on the perturbation magnitude and is insufficient to directly establish that the perturbed gradients preserve the information necessary for effective model optimization. Consequently, the current theoretical analysis provides limited insight into the practical trade-off between label privacy and training utility, calling for further theoretical justification or experimental validation.
Questions for authors
  1. Robustness Against Adaptive Attacks. Gradient Mirage disrupts the gradient-label correspondence through objective, directional, and scale perturbations. However, could an attacker aware of the defense mechanism recover private labels by jointly estimating supervision masks, token-wise scaling factors, and directional perturbations? Although the appendix considers a scale-aware adaptive attack, it only optimizes a global scaling parameter. Could the authors further justify or experimentally evaluate the robustness against stronger adaptive attacks?
  2. Connection Between Theoretical Privacy Guarantees and Practical Label Protection. The vMF mechanism provides Metric Differential Privacy guarantees over gradient directions, but how do these guarantees translate into privacy protection for the original text or label sequences? Given the relatively large privacy budgets (ε = 512 or 1024), could the authors clarify their practical privacy implications and their relationship to the low reconstruction scores observed experimentally?
  3. Training Utility Across Different Models. Table 1 shows noticeable variations in training utility across models, particularly the increased PPL of DeepSeek-LLM-7B compared with standard training. How do the authors explain these differences? Are they primarily caused by selective supervision, scale blinding, or directional perturbation? Could the authors clarify the conditions under which Gradient Mirage maintains a favorable privacy–utility trade-off across different models?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 1zVL (Part 1/2)

Dear Reviewer,

We sincerely thank you for the thoughtful and valuable questions and suggestions. We truly appreciate the careful attention you have given to our work, and we believe your comments reflect a thorough understanding of our paper. In the following, we will address your concerns in detail and provide additional explanations, experiments, and improvements based on your valuable suggestions.

W1/Q1: Limited Robustness Evaluation Against Sophisticated Adaptive Attacks and Multiple Gradient Observations

We sincerely thank you for raising this important concern.

In fact, our SAS design is intended, to some extent, to mitigate attacks that aggregate multiple gradient observations. First, due to the random token sampling controlled by ρ\rho, the supervised token subsets across different training steps may overlap considerably. Therefore, aggregating multiple observations does not necessarily recover the complete supervision information of the original sequence, as a substantial amount of redundancy may exist among the selected tokens. Second, gradients observed at different training steps are generated under different model parameter snapshots, which introduces additional inconsistencies and makes direct gradient aggregation more challenging. Finally, an attacker needs to determine whether multiple gradient observations correspond to the same original sequence before aggregating them. Without explicit sample identifiers, sequence length alone is insufficient for reliable identification, since different training samples may share the same length.

To further address the concern about more sophisticated adaptive attacks, we have supplemented our evaluation with a stronger adaptive attack. Specifically, we introduce an independently optimizable weight for each token during gradient matching, allowing the attacker to jointly optimize the reconstructed label sequence and token-wise scaling factors. This setting enables the attacker to adaptively compensate for token-level scaling perturbations, making it more challenging than the global-scale adaptive attack considered in our original appendix. The corresponding results are presented in the main text (Table 2), and we also include them here:

ModelεToken-wise ROUGE-L F1 ↓BiSR(b) ROUGE-L F1 ↓Δ ROUGE-L F1 (pp)Token-wise TRR (%) ↓BiSR(b) TRR (%) ↓Δ TRR (pp)
Llama3-8B128.05 ± .02.04+1.147.65 ± 2.085.00+2.65
256.10 ± .02.06+3.529.27 ± 2.127.00+2.27
384.13 ± .03.09+3.5012.17 ± 2.149.00+3.17
512.23 ± .03.20+2.8427.77 ± 2.2425.00+2.77
1024.36 ± .03.35+1.0239.91 ± 2.5537.00+2.91
Qwen2.5-14B128.15 ± .04.13+1.996.39 ± 7.026.00+0.39
256.19 ± .05.15+4.3610.44 ± 7.968.00+2.44
384.26 ± .05.18+7.6117.08 ± 7.7611.00+6.08
512.30 ± .03.24+5.6924.31 ± 5.1122.00+2.31
1024.34 ± .03.29+5.0933.03 ± 4.7826.00+7.03
AuthorsOpenReview ↗

Response to Reviewer 1zVL (Part 2/2)

These results suggest that the combination of vMF-based directional perturbations and parameter drift during model training makes it substantially more challenging for the attacker to jointly optimize the token-wise weight mask and the reconstructed label sequence. Although an attacker could alternatively attempt to infer the supervision pattern by exhaustively searching over possible token positions, such a strategy would incur substantial computational overhead. For a sequence of length LL, there are up to 2L2^L possible binary token-selection masks, making exhaustive enumeration computationally infeasible for typical sequence lengths. Even a position-wise search would require repeated gradient computations and reconstruction attempts, with no guarantee of identifying the correct supervision pattern under noisy gradient observations. Furthermore, the continuous scaling factors introduced by our Scale Blinding mechanism further increase the complexity of exhaustive search.

W2/Q2: Relationship of the vMF Privacy Budget ε\varepsilon to Sequence-Level Privacy and Fine-Tuning Utility

We frankly acknowledge that we have not yet established a unified parameter that directly characterizes the privacy–utility trade-off at the sequence level. Nevertheless, we have made every effort to provide practical guidance on selecting the privacy budget ε\varepsilon of the vMF mechanism (Directional MDP) through experimental evaluations.

We observed strong defense effectiveness across all evaluated models under the setting of ε = 512. Additionally, we evaluated the defense effectiveness and utility under smaller privacy budgets on Llama3-8B and Qwen2.5-14B. These results suggest that privacy budgets below ε = 512 can provide effective privacy protection, while still allowing some flexibility in the privacy–utility trade-off.

ModelMetric128192256384
Llama3-8BPPL/Acc(%)2.546/43.292.523/38.412.513/35.372.498/44.51
RougeL-F0.040.040.060.09
Qwen2.5-14BPPL/Acc(%)2.372/68.292.366/65.242.357/65.852.349/65.24
RougeL-F0.130.140.150.18

Since the gradients at the split interface provide the only signal associated with the labels, the indistinguishability of their directions is closely related to label indistinguishability. As our work primarily focuses on an empirical approach to privacy protection, we report empirical measurements here.

Q3: Impact of Defense Components on Model Utility

We sincerely thank you for raising this important question.

Gradient Mirage is fundamentally a privacy-preserving fine-tuning method that trades a certain degree of training utility for stronger privacy protection, rather than achieving completely lossless privacy preservation. All three defense components may introduce some degree of optimization disturbance.

The impact of these perturbations can vary across models due to differences in pretrained parameters, model architectures, and optimization characteristics. This may partly explain why DeepSeek-LLM-7B exhibits a more noticeable PPL increase compared with standard training. Since PPL alone may not fully capture differences in model utility, we have supplemented Table 1 with Task-Level Utility Evaluation results. Furthermore, we have extended our experiments to three additional larger-scale models, including Qwen3-30B-A3B, which adopts a Mixture-of-Experts (MoE) architecture. The results demonstrate that, under our recommended hyperparameter settings, Gradient Mirage achieves both effective privacy protection and competitive downstream task utility across these models.

Closing Remarks

We sincerely appreciate your thoughtful and exceptionally insightful review. Your questions demonstrate a deep understanding of our work and highlight important challenges concerning adaptive attacks, theoretical privacy guarantees, and practical utility. We are particularly grateful for your positive evaluation and constructive suggestions, which have helped us further strengthen our manuscript.

We hope that our additional experiments and detailed clarifications have adequately addressed your concerns. We would be deeply grateful if you could consider further improving your already positive assessment in light of these revisions. Thank you once again for your time, encouragement, and valuable feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary of Contributions and Rebuttal Improvements

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we would like to take this opportunity to highlight the key contributions of our work and the additional improvements we have made based on your valuable feedback.

Our work aims to address a critical privacy vulnerability in LLM Split Learning (LLM-SL): gradient-based label leakage. Existing forward-pass defenses, such as ObfusLM and Dp-Forward, can disrupt forward inversion attacks through embedding obfuscation or noise injection. However, they provide limited protection against backward gradient leakage, since GMA-SL can directly use the perturbed hidden states xtrkx_{\mathrm{trk}} to compute gradients within its reconstruction graph. Gradient Mirage aims to close this gap by disrupting the correspondence between exposed gradients and private labels.

In response to the reviewers' valuable suggestions, we have conducted extensive additional experiments and made several improvements to our manuscript. Specifically, we evaluated our defense against stronger adaptive attacks, including token-wise adaptive gradient matching, and extended our experiments to larger LLMs and an MoE-based LLM. We also investigated training stability and utility under smaller directional MDP privacy budgets, supplemented downstream task-level utility evaluations, and provided a clearer explanation of our threat model. These additional experiments and clarifications have been incorporated into our revised manuscript, and we have carefully addressed the reviewers' follow-up questions.

We sincerely thank the Area Chair and all reviewers for your time, careful consideration, and constructive feedback. Your suggestions have helped us strengthen both the experimental evaluation and the presentation of our work. We hope that our additional results and explanations have adequately addressed your concerns. If so, we would be deeply grateful if the reviewers could kindly reconsider their ratings and the Area Chair could take these updates into account in the final assessment.

Once again, we sincerely appreciate your time, thoughtful reviews, and valuable suggestions.

Best regards,

The Authors of Gradient Mirage

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.