Pith. sign in

REVIEW 3 major objections 6 minor 50 references

FedQSN claims that masking and quantizing model parameters keeps the server's final model accurate while giving clients a measurably weaker proxy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

FedQSN hides part of the server model with random masks and quantizes the remainder to give clients a degraded proxy, reporting a large global-vs-proxy performance gap with modest loss in the final global model.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A plausible masking+quantization recipe for FL model privacy, but Algorithm 1 never re-applies the server mask, so the central confidentiality claim is unsupported as written. the 3 major comments →

arxiv 2508.18911 v2 pith:MLHTCD3V submitted 2025-08-26 cs.LG cs.AI

Enhancing Model Privacy in Federated Learning with Random Masking and Quantization

Zhibo Xu , Jianhao Zhu , Jingwen Xu , Changze Lv , Zisu Huang , Xiaohua Wang , Muling Wu , Qi Qian
show 2 more authors
Xiaoqing Zheng Xuanjing Huang
This is my paper
classification cs.LG cs.AI
keywords federated learningmodel privacyintellectual property protectionrandom maskingquantizationlarge language modelsproxy modelsparameter confidentiality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

FedQSN is a federated-learning method aimed at protecting the server's model as well as clients' data, in settings where a company fine-tunes a proprietary large language model across remote clients. It sends each client a proxy model made by zeroing a random subnetwork (a server mask fixed across clients plus a client-specific mask) and quantizing the remaining weights to low precision. The paper's claim is that this proxy keeps the final global model competitive with FedAvg while giving the server a large, measurable performance advantage over what any client can see. That advantage—quantified by the gap between global and proxy model scores and by low parameter cosine similarity—is offered as evidence that clients cannot reverse-engineer the server's proprietary model. The method matters because existing approaches either require the server to hold an IID auxiliary dataset or leave the model backbone exposed.

Core claim

FedQSN's central claim is that a server can train a high-quality global model in federated learning without ever showing a complete, exact set of weights to any client. The server applies a server mask that blanks a fixed subset of parameters for all clients, then for each selected client applies a fresh client-specific mask, then quantizes the surviving weights into low-bit blocks; the result is a privacy-preserving proxy model. Clients train this proxy locally and return updates; the server aggregates them into the global model. At the end, a logical AND between the trained model and the original untrained model restores the masked parameters. The paper argues this yields two things at onc

What carries the argument

The load-bearing mechanism is a two-level random-masking scheme combined with blockwise quantization. A column-wise mask zeroes whole parameter columns with probability p and rescales survivors by 1/(1-p), following Dropout's expectation-preserving trick, so the masked proxy has the same expected output magnitude as the real model. The server mask is meant to hide the same subnetwork from every client, blocking coalition-based reconstruction; the client mask makes each client train a different subnetwork each round so all parameters still get updated over time; and omega-bit block quantization turns exact floats into coarse integers, which both lowers communication cost and prevents exact re

Load-bearing premise

The central privacy claim assumes the server mask hides a fixed subset of parameters from every client in every round; if instead the mask is applied only to the initial model, the hidden parameters become visible to clients in later rounds and the no-reconstruction guarantee collapses.

What would settle it

Inspect the round-2 client proxies produced by Algorithm 1: if any parameter zeroed by the server mask at W0 is nonzero in the distributed W1, then the server mask is not re-applied and a client coalition can see previously hidden parameters. A second check: train an attacker that averages the client proxies from a round and see whether the result matches the server's global model more closely than the reported cosine-similarity metric suggests.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Servers can fine-tune proprietary LLMs with remote clients while keeping a measurable edge: the proxy model's performance is consistently below the global model's across BLEU, NIST, METEOR, ROUGE-L, and CIDEr.
  • FedQSN removes the need for an IID auxiliary dataset on the server, so model protection does not come at the cost of the server seeing data that resembles client data.
  • The privacy-performance trade-off is tunable through quantization bit-width and mask ratios; the reported settings around 3-4 bits with modest masking keep global quality high while widening the proxy gap.
  • The method is demonstrated across model scales from GPT-2 to Llama3.1-8B and across cross-silo and cross-device FL scenarios, so the claimed protection is not tied to one architecture.
  • The authors position FedQSN as orthogonal to data-level defenses, meaning it can be layered with differential-privacy or noise-based methods rather than replacing them.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A reader following Algorithm 1 will notice that the server mask is applied only to W0 at line 3, while the prose promises a mask on the model distributed each round; the strongest privacy claim therefore implicitly requires the server mask to be re-applied to each aggregated W_t, and that gap should be resolved by a direct implementation check.
  • The global-vs-proxy performance gap and cosine-similarity metric are coarse proxies for IP leakage; a sharper check would be a reconstruction attack that inverts several proxy models, or membership-inference on proprietary training data, which FedQSN does not address.
  • A testable extension is to re-mask the aggregated model every round and measure whether the proxy gap widens without hurting convergence; if it does, the method's privacy could be hardened without changing its architecture.
  • The reported results show that very aggressive quantization (1-2 bits) hurts global-model utility, so the practical operating region is narrow; an adaptive bit-width schedule—coarse early in training, finer later—might preserve privacy while closing that gap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes FedQSN, a federated learning method intended to protect server-side model IP. The server applies a random mask to a subset of model parameters (server mask) and then a client-specific mask (client mask), followed by blockwise quantization, before sending a 'proxy model' to clients. Clients fine-tune the proxy and return full updates, which the server aggregates. The paper claims that the server mask prevents any client coalition from reconstructing the complete model, and that final logical AND with the original untrained model restores masked parameters. Experiments on GPT-2 and Llama models over four NLG datasets compare FedQSN with FedAvg, FedSP, and FedLPP, using the performance gap between the global and proxy models and parameter cosine similarity as privacy metrics. The global model is reported to stay close to FedAvg while the proxy gap is larger than for FedLPP.

Significance. If the central confidentiality claim were valid, the paper would make a useful contribution to model-IP protection in federated learning: it is simple, does not require server-side labeled data, and is compatible with secure aggregation. The authors provide code, cover multiple model scales, and include ablations of the mask ratios and quantization bit widths. However, the load-bearing privacy claim is contradicted by the formal algorithm as written, and the privacy evaluation does not measure adversarial confidentiality. The contribution as stated is therefore not established.

major comments (3)
  1. [Algorithm 1, lines 3 and 7; Section 3.4] The server mask is applied only to the initial model W0 (line 3). Inside the training loop each client receives W_{t-1} masked only with the client-specific p2 mask (line 7). Clients return full fine-tuned models W^c_t, and the server aggregates them into W_t (line 11). Positions that were zeroed by the server mask at round 1 receive nonzero local updates and therefore appear in W_t. From round 2 onward clients observe those positions in the distributed model, up to the per-client p2 mask. A single client observing multiple rounds, or a coalition, can thus see the parameters that Section 3.4 claims are hidden. The prose-level statement that the server mask 'prevents any client coalition from reconstructing the complete model' is not supported by the pseudocode; the mask must be reapplied every round for that claim to hold, and the algorithm does not do so.
  2. [Section 3.4 and Figure 1 caption; Eq. (2)] The 'logical AND operation' between the final aggregated model and the original untrained model is formally undefined and internally inconsistent. The mask entries in Eq. (2) are non-binary (0 or 1/(1-p)), so 'logical AND' has no clear meaning. If the server mask is not reapplied, there are no masked coordinates left to restore. If it is reapplied, ANDing the final aggregated model with the original untrained model would reset the hidden coordinates to their initialization values, so the result would not be a 'fully trained global model' as claimed. The final reconstruction step needs a precise definition and a proof that it preserves the trained values of the hidden parameters.
  3. [Section 4.2 and Table 6] The privacy evaluation does not measure the claimed confidentiality property. The performance gap between the global model and the proxy, and the cosine similarity between their parameters, are direct consequences of applying masking and quantization: a larger gap only shows that the proxy is more distorted. It does not demonstrate that a client or coalition cannot recover the server's model parameters or functionality, which is the paper's stated guarantee. The protocol is inherited from FedLPP, and comparing FedQSN with FedLPP under the same metric is circular for establishing an absolute level of privacy. The paper needs an explicit adversary model and an evaluation that tests reconstruction or inference, rather than treating utility distortion as a privacy measure.
minor comments (6)
  1. [Algorithm 1] Function signature inconsistency: the call is ClientLocalTrain(W^c_t, c) on lines 7-9, but the definition is ClientLocalTrain(c, Wc) on line 14. Also, the pseudocode does not show the final logical AND step described in Section 3.4.
  2. [Eq. (2) and Section 3.2] The text says the mask is applied 'column-wise,' but Eq. (2) describes independently sampled entries. These are different masking operations and should be disambiguated.
  3. [Section 3.3] The claim that quantization 'hinders accurate reconstruction' and 'obscures exact parameter values' is asserted without support. Blockwise absmax quantization is deterministic and approximately invertible if the block size and bit width are known; the paper should either provide an attack analysis or moderate the claim.
  4. [Table 3] Typo: 'porxy' should be 'proxy.'
  5. [References] The BLEU reference is listed as 'Papinesi, 2002'; the correct spelling is Papineni.
  6. [Figure 2] The subplots are very small and the axis labels are difficult to read. A higher-resolution figure with clearer grouping would be helpful.

Circularity Check

2 steps flagged

Privacy 'enhancement' is measured as the proxy-vs-global gap, which is exactly the quantity produced by the masking and quantization being evaluated; the metric and baseline are inherited from the authors' own FedLPP paper.

specific steps
  1. self definitional [Section 4.2, 'Performance Comparison' (evaluation protocol); cf. Eq. 3 and Eq. 4]
    "To evaluate server-side model privacy protection, we follow the evaluation protocol proposed in FedLPP (Zhu et al., 2024), which uses the performance gap between the server-held global model and the client-received proxy model as a measure of privacy preservation. ... Moreover, a larger performance gap suggests that the proxy model remains less competitive, providing stronger evidence of privacy preservation."

    Privacy is operationalized as the performance gap between the server model and the proxy. The proxy is constructed from the same server model by the very operations being evaluated: random masking (Eq. 3) and quantization (Eq. 4). The gap is therefore a direct, tunable function of p1, p2, and omega, not an independent measurement. Reporting that FedQSN 'enhances privacy' because the gap is larger is equivalent to reporting that FedQSN degrades its proxy more; the output metric is the input mechanism. The evaluation protocol is itself imported, by self-citation, from the authors' prior FedLPP work, so no external privacy test is introduced.

  2. self citation load bearing [Section 1 and Section 4.2 (FedLPP baseline and evaluation protocol)]
    "FedLPP (Zhu et al., 2024) eliminates the need for such auxiliary datasets by sharing only quantized LoRA (Low-Rank Adaptation) adapters with clients, thus enhancing model privacy. ... To evaluate server-side model privacy protection, we follow the evaluation protocol proposed in FedLPP (Zhu et al., 2024), which uses the performance gap between the server-held global model and the client-received proxy model as a measure of privacy preservation."

    FedLPP is the same group's prior work (shared co-authors: Jianhao Zhu, Changze Lv, Xiaohua Wang, Muling Wu, Xiaoqing Zheng, Xuanjing Huang). The paper adopts FedLPP's own evaluation protocol as the primary definition of model privacy, then claims its method beats FedLPP under that definition. The privacy axis of the comparison is therefore supplied by a self-citation rather than by an external privacy definition, reconstruction attack, or formal guarantee. The citation is load-bearing: without this protocol there is no independent measure of 'parameter-level confidentiality.'

full rationale

The circularity is concentrated in the privacy half of the paper. Section 4.2 explicitly borrows FedLPP's protocol and defines 'privacy preservation' as the global-vs-proxy performance gap. Because masking (Eq. 3) and quantization (Eq. 4) are exactly the operations used to construct the proxy from the global model, the reported gap (e.g., 8.76 vs 2.14 in Table 1) is a direct, tunable consequence of the method, not an independent measurement. Larger masking ratios and lower bit-widths produce larger gaps by construction, as the ablation tables confirm; calling this 'enhanced parameter-level confidentiality' partially restates the mechanism as its own evidence. The metric and the main baseline FedLPP also originate from the same research group's prior paper, so the self-citation is load-bearing for the central privacy claim. On the other hand, the performance comparisons against FedAvg (global-model utility) are genuine external measurements and show that the method does not simply collapse; that part is not circular. A separate, serious correctness issue--not counted as circularity--is that Algorithm 1 applies the server mask only to W0 (line 3) and inside the training loop applies only the p2 client mask (line 7), so the claimed 'no client coalition can reconstruct the complete model' guarantee does not follow from the stated protocol. Score 6 reflects partial circularity in the central privacy claim.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method's operation depends on hand-chosen mask ratios p1, p2 and bit width omega, all tuned per dataset. It also assumes the standard FL aggregation objective and the expectation-preserving property of the mask. No new physical entities are introduced.

free parameters (5)
  • server mask ratio p1 = 0.1 (searched over {0.05, 0.1, 0.15, 0.2})
    Controls the fraction of weights hidden from all clients; chosen per dataset and affects privacy and performance.
  • client mask ratio p2 = 0.1 (searched over {0.05, 0.1, 0.15, 0.2})
    Adds per-client masking; selected by hyperparameter search.
  • quantization bit width omega = 2 or 3-4 in main results (searched over {1,2,3,4})
    Sets precision of proxy model; the paper presents a privacy-performance tradeoff scan and picks values that balance them.
  • quantization block size s = 256
    Fixed block size for absmax quantization; affects granularity and proxy fidelity.
  • communication rounds R = 10
    Standard in their setup, but short for LLM fine-tuning.
axioms (4)
  • standard math Federated averaging objective (Eq. 1) correctly represents the FL setup.
    Used to define the learning problem.
  • standard math Mask expectation E[W_tilde]=W holds so that masking does not shift activations.
    Section 3.2, depends on independent Bernoulli entries.
  • domain assumption The performance gap between global and proxy models is a valid measure of model privacy.
    Adopted from FedLPP; no independent validation that gap correlates with resistance to model extraction or inversion.
  • domain assumption Server mask persists across rounds to keep hidden parameters secret.
    Implicit in the privacy claim, but Algorithm 1 applies it only to W0.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Model Privacy in Federated Learning with Random Masking and Quantization." pith.science (2026). https://pith.science/paper/MLHTCD3V

@misc{pith2026250818911,
  author       = {Pith},
  title        = {Pith review of: Enhancing Model Privacy in Federated Learning with Random Masking and Quantization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLHTCD3V}},
  note         = {Machine review of arXiv:2508.18911}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The primary goal of traditional federated learning is to protect data privacy by enabling distributed edge devices to collaboratively train a shared global model while keeping raw data decentralized at local clients. The rise of large language models (LLMs) has introduced new challenges in distributed systems, as their substantial computational requirements and the need for specialized expertise raise critical concerns about protecting intellectual property (IP). This highlights the need for a federated learning approach that can safeguard both sensitive data and proprietary models. To tackle this challenge, we propose FedQSN, a federated learning approach that leverages random masking to obscure a subnetwork of model parameters and applies quantization to the remaining parameters. Consequently, the server transmits only a privacy-preserving proxy of the global model to clients during each communication round, thus enhancing the model's confidentiality. Experimental results across various models and tasks demonstrate that our approach not only maintains strong model performance in federated learning settings but also achieves enhanced protection of model parameters compared to baseline methods.

Figures

Figures reproduced from arXiv: 2508.18911 by Changze Lv, Jianhao Zhu, Jingwen Xu, Muling Wu, Qi Qian, Xiaohua Wang, Xiaoqing Zheng, Xuanjing Huang, Zhibo Xu, Zisu Huang.

Figure 1
Figure 1. Figure 1: The proposed federated learning approach protects both server-side model and client-side data privacy. (1) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of global and proxy models across multiple metrics for FedQSN on various datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65--72

    2005

  2. [2]

    Rouzbeh Behnia, Mohammadreza Reza Ebrahimi, Jason Pacheco, and Balaji Padmanabhan. 2022. Ew-tune: A framework for privately fine-tuning large language models with differential privacy. In 2022 IEEE International Conference on Data Mining Workshops (ICDMW), pages 560--566. IEEE

    work page 2022

  3. [3]

    Anja Belz and Ehud Reiter. 2006. Comparing automatic and human evaluation of nlg systems

    work page 2006

  4. [4]

    Dan Bogdanov, Sven Laur, and Jan Willemson. 2008. Sharemind: A framework for fast privacy-preserving computations. In Computer Security-ESORICS 2008: 13th European Symposium on Research in Computer Security, M \'a laga, Spain, October 6-8, 2008. Proceedings 13 , pages 192--206. Springer

    work page 2008

  5. [5]

    Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2017. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175--1191

    work page 2017

  6. [6]

    Xingjuan Cai, Yun Niu, Shaojin Geng, Jiangjiang Zhang, Zhihua Cui, Jianwei Li, and Jinjun Chen. 2020. An under-sampled software defect prediction method based on hybrid multi-objective cuckoo search. Concurrency and Computation: Practice and Experience, 32(5):e5478

    work page 2020

  7. [7]

    Zachary Charles, Arun Ganesh, Ryan McKenna, H Brendan McMahan, Nicole Mitchell, Krishna Pillutla, and Keith Rush. 2024. Fine-tuning large language models with user-level differential privacy. arXiv preprint arXiv:2407.07737

    Pith/arXiv arXiv 2024

  8. [8]

    Yulong Chen, Yang Liu, Liang Chen, and Yue Zhang. 2021. Dialogsum: A real-life scenario dialogue summarization dataset. arXiv preprint arXiv:2105.06762

    Pith/arXiv arXiv 2021

  9. [9]

    Lin Chin-Yew. 2004. Rouge: A package for automatic evaluation of summaries. In Proceedings of the Workshop on Text Summarization Branches Out, 2004

    work page 2004

  10. [10]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1--113

    2023

  11. [11]

    Maurizio Colombo, Rasool Asal, Ernesto Damiani, Lamees M AlQassem, Al Anoud Almemari, and Yousof Alhammadi. 2025. A quantization-based technique for privacy preserving distributed learning. Future Generation Computer Systems, page 107741

    work page 2025

  12. [12]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36

    2024

  13. [13]

    Chenhe Dong, Yuexiang Xie, Bolin Ding, Ying Shen, and Yaliang Li. 2023. Tunable soft prompts are messengers in federated learning. arXiv preprint arXiv:2311.06805

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Robin C Geyer, Tassilo Klein, and Moin Nabi. 2017. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557

    Pith/arXiv arXiv 2017

  15. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv--2407

    2024

  16. [16]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790--2799. PMLR

    2019

  17. [17]

    Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. 2021. A survey on federated learning for resource-constrained iot devices. IEEE Internet of Things Journal, 9(1):1--24

    work page 2021

  18. [18]

    Eunjeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. 2018. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. arXiv preprint arXiv:1811.11479

    Pith/arXiv arXiv 2018

  19. [19]

    Weizhao Jin, Yuhang Yao, Shanshan Han, Jiajun Gu, Carlee Joe-Wong, Srivatsan Ravi, Salman Avestimehr, and Chaoyang He. 2023. Fedml-he: An efficient homomorphic-encryption-based privacy-preserving federated learning system. arXiv preprint arXiv:2303.10837

    Pith/arXiv arXiv 2023

  20. [20]

    Juraj Juraska, Kevin K Bowden, and Marilyn Walker. 2019. Viggo: A video game corpus for data-to-text generation in open-domain conversation. arXiv preprint arXiv:1910.12129

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  21. [21]

    Peter Kairouz, H Brendan McMahan, Brendan Avent, Aur \'e lien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. 2021. Advances and open problems in federated learning. Foundations and trends in machine learning , 14(1--2):1--210

    work page 2021

  22. [23]

    Jakub Konecn \`y , H Brendan McMahan, Felix X Yu, Peter Richt \'a rik, Ananda Theertha Suresh, and Dave Bacon. 2016. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 8

    Pith/arXiv arXiv 2016

  23. [24]

    Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xiangfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. 2022. Fedscale: Benchmarking model and system performance of federated learning at scale. In International conference on machine learning, pages 11814--11827. PMLR

    work page 2022

  24. [25]

    Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. 2019. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189

    Pith/arXiv arXiv 2019

  25. [26]

    Wei Yang Bryan Lim, Nguyen Cong Luong, Dinh Thai Hoang, Yutao Jiao, Ying-Chang Liang, Qiang Yang, Dusit Niyato, and Chunyan Miao. 2020. Federated learning in mobile edge networks: A comprehensive survey. IEEE communications surveys & tutorials, 22(3):2031--2063

    work page 2020

  26. [27]

    Guodong Long, Yue Tan, Jing Jiang, and Chengqi Zhang. 2020. Federated learning for open banking. In Federated Learning: Privacy and Incentive, pages 240--254. Springer

    work page 2020

  27. [28]

    Wang Luping, WANG Wei, and LI Bo. 2019. Cmfl: Mitigating communication overhead for federated learning. In 2019 IEEE 39th international conference on distributed computing systems (ICDCS), pages 954--964. IEEE

    work page 2019

  28. [29]

    Peihua Mai, Ran Yan, Zhe Huang, Youjia Yang, and Yan Pang. 2023. Split-and-denoise: Protect large language model inference with local differential privacy. arXiv preprint arXiv:2310.09130

    Pith/arXiv arXiv 2023

  29. [30]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273--1282. PMLR

    2017

  30. [31]

    Linyong Nan, Dragomir Radev, Rui Zhang, Amrit Rau, Abhinand Sivaprasad, Chiachun Hsieh, Xiangru Tang, Aadit Vyas, Neha Verma, Pranav Krishna, et al. 2020. Dart: Open-domain structured data record to text generation. arXiv preprint arXiv:2007.02871

    Pith/arXiv arXiv 2020

  31. [32]

    Dinh C Nguyen, Quoc-Viet Pham, Pubudu N Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia Dobre, and Won-Joo Hwang. 2022. Federated learning for smart healthcare: A survey. ACM Computing Surveys (Csur), 55(3):1--37

    work page 2022

  32. [33]

    Jekaterina Novikova, Ond r ej Du s ek, and Verena Rieser. 2017. The e2e dataset: New challenges for end-to-end generation. arXiv preprint arXiv:1706.09254

    Pith/arXiv arXiv 2017

  33. [34]

    K Papinesi. 2002. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311--318

    work page 2002

  34. [35]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  35. [36]

    Parizi, Quan Z

    Osama Shahid, Seyedamin Pouriyeh, Reza M. Parizi, Quan Z. Sheng, Gautam Srivastava, and Liang Zhao. 2021. https://arxiv.org/abs/2107.10996 Communication efficiency in federated learning: Achievements and challenges . Preprint, arXiv:2107.10996

    Pith/arXiv arXiv 2021

  36. [37]

    Tanmay Singh, Harshvardhan Aditya, Vijay K Madisetti, and Arshdeep Bahga. 2024. Whispered tuning: Data privacy preservation in fine-tuning llms through differential privacy. Journal of Software Engineering and Applications, 17(1):1--22

    work page 2024

  37. [38]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929--1958

    work page 2014

  38. [39]

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566--4575

    2015

  39. [40]

    Rui Wang, Tong Yu, Ruiyi Zhang, Sungchul Kim, Ryan Rossi, Handong Zhao, Junda Wu, Subrata Mitra, Lina Yao, and Ricardo Henao. 2024. Personalized federated learning for text classification with gradient-free prompt tuning. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4597--4612

    work page 2024

  40. [41]

    Thomas Wolf. 2020. Transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771

    Pith/arXiv arXiv 2020

  41. [42]

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087--38099. PMLR

    2023

  42. [43]

    Jing Xu and Jingzhao Zhang. 2024. Random masking finds winning tickets for parameter efficient fine-tuning. arXiv preprint arXiv:2405.02596

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [44]

    Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. 2019. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1--19

    work page 2019

  44. [45]

    Zhewei Yao, Cheng Li, Xiaoxia Wu, Stephen Youn, and Yuxiong He. 2023. A comprehensive study on post-training quantization for large language models. arXiv preprint arXiv:2303.08302

    Pith/arXiv arXiv 2023

  45. [46]

    Zeling Zhang, Dongqi Cai, Yiran Zhang, Mengwei Xu, Shangguang Wang, and Ao Zhou. 2024. Fedrdma: Communication-efficient cross-silo federated llm via chunked rdma transmission. In Proceedings of the 4th Workshop on Machine Learning and Systems, pages 126--133

    work page 2024

  46. [47]

    Zhuo Zhang, Yuanhang Yang, Yong Dai, Qifan Wang, Yue Yu, Lizhen Qu, and Zenglin Xu. 2023. Fedpetuning: When federated learning meets the parameter-efficient tuning methods of pre-trained language models. In Annual Meeting of the Association of Computational Linguistics 2023, pages 9963--9977. Association for Computational Linguistics (ACL)

    work page 2023

  47. [48]

    Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. 2018. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582

    Pith/arXiv arXiv 2018

  48. [49]

    JianHao Zhu, Changze Lv, Xiaohua Wang, Muling Wu, Wenhao Liu, Tianlong Li, Zixuan Ling, Cenyuan Zhang, Xiaoqing Zheng, and Xuanjing Huang. 2024. Promoting data and model privacy in federated learning through quantized lora. arXiv preprint arXiv:2406.10976

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [50]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  50. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.