When Differential Privacy Meets Wireless Federated Learning: An Improved Analysis for Privacy and Convergence

Chen Yaoling; Liang Hao; Tu Xiaotong

arxiv: 2603.19040 · v1 · submitted 2026-03-19 · 💻 cs.LG

When Differential Privacy Meets Wireless Federated Learning: An Improved Analysis for Privacy and Convergence

Chen Yaoling , Liang Hao , Tu Xiaotong This is my paper

Pith reviewed 2026-05-15 08:08 UTC · model grok-4.3

classification 💻 cs.LG

keywords differential privacyfederated learningwireless networksnon-convex optimizationconvergence analysisprivacy accountinggradient clipping

0 comments

The pith

In differentially private wireless federated learning, the total privacy loss converges to a fixed constant instead of growing without bound over many iterations.

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

The paper analyzes privacy and convergence for wireless federated learning under differential privacy when the loss function is smooth and non-convex. It incorporates the effects of random device selection and mini-batch sampling into the privacy accounting and shows that cumulative privacy loss reaches a plateau. The analysis also supplies convergence guarantees that include gradient clipping and produces an explicit trade-off between the privacy budget and final model utility. A reader would care because these results remove a practical obstacle to running private learning systems over wireless links where participation fluctuates and channel noise is unavoidable.

Core claim

For differentially private wireless federated learning with general smooth non-convex objectives, explicit modeling of device selection and mini-batch sampling shows that privacy loss converges to a constant rather than diverging with the number of iterations; convergence guarantees hold under gradient clipping and an explicit privacy-utility trade-off is derived.

What carries the argument

The privacy accounting procedure that tracks cumulative loss under random device participation and mini-batch sampling, which produces a bounded total privacy loss.

Load-bearing premise

The analysis assumes that wireless channel conditions, device participation patterns, and noise addition follow the specific probabilistic models used in the derivation.

What would settle it

Track the measured privacy loss over several hundred communication rounds in a wireless federated learning experiment; if the loss keeps increasing roughly linearly with rounds instead of flattening to a constant, the central claim does not hold.

read the original abstract

Differentially private wireless federated learning (DPWFL) is a promising framework for protecting sensitive user data. However, foundational questions on how to precisely characterize privacy loss remain open, and existing work is further limited by convergence analyses that rely on restrictive convexity assumptions or ignore the effect of gradient clipping. To overcome these issues, we present a comprehensive analysis of privacy and convergence for DPWFL with general smooth non-convex loss objectives. Our analysis explicitly incorporates both device selection and mini-batch sampling, and shows that the privacy loss can converge to a constant rather than diverge with the number of iterations. Moreover, we establish convergence guarantees with gradient clipping and derive an explicit privacy-utility trade-off. Numerical results validate our theoretical findings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper bounds privacy loss to a constant in wireless DPFL by folding in device selection and channel noise, while extending convergence to non-convex objectives with clipping.

read the letter

The punchline is that this work shows privacy loss in differentially private wireless federated learning can be bounded by a constant rather than growing without bound as the number of rounds increases. They achieve this through a refined analysis that accounts for random device selection and the noise introduced by wireless channels, while also providing convergence results for non-convex objectives under gradient clipping. What stands out is how they integrate device selection and mini-batch sampling directly into the privacy accounting. This is not just a minor tweak; it allows the per-round privacy leakage to be controlled in a way that the total loss plateaus. Extending the analysis beyond convex losses is helpful, as many real models are non-convex. They also derive a clear privacy-utility trade-off and include numerical validation that matches the theory. On the downside, the constant privacy bound is quite dependent on the specific assumptions in their system model, particularly the independence of device participation and the variance of the wireless channel noise. If those conditions are not met in practice, such as with correlated device behaviors or different fading models, the privacy loss would likely resume its growth according to standard composition results. The convergence guarantees are reasonable but follow expected patterns for clipped non-convex optimization, so the novelty there is more in the combination with the privacy part. This paper targets researchers focused on federated learning in wireless environments who need to balance privacy with performance. The approach shows clear thinking about the interactions between communication constraints and privacy mechanisms. Given the theoretical contributions and the practical relevance, it should go to peer review rather than being desk rejected.

Referee Report

2 major / 2 minor

Summary. The paper analyzes differentially private wireless federated learning (DPWFL) under general smooth non-convex loss functions. It incorporates device selection and mini-batch sampling into the privacy accounting to show that privacy loss converges to a constant (rather than diverging with iterations T), establishes convergence rates with gradient clipping, and derives an explicit privacy-utility trade-off. Numerical results are presented to support the claims.

Significance. If the derivations hold, the work would meaningfully advance DP-FL analysis by relaxing convexity assumptions and folding realistic wireless channel noise plus random participation into the privacy bound, yielding a non-diverging privacy loss that is not available from standard advanced composition. This could improve practical privacy-utility tuning in wireless settings, provided the modeled selection and noise assumptions are representative.

major comments (2)

[Privacy analysis section] Privacy analysis (around the composition theorem and Theorem on privacy loss): the headline claim that privacy loss converges to a constant rather than diverging with T rests on folding independent device selection probability p and additive wireless channel noise variance sigma_w^2 into the per-round leakage. The derivation should explicitly bound the sensitivity under these models and state the precise conditions (e.g., independence of selections across rounds) under which the constant bound holds; otherwise standard composition recovers linear or sqrt(T) growth.
[Convergence guarantees section] Convergence analysis (section on non-convex rates with clipping): the guarantees for smooth non-convex objectives with gradient clipping require an explicit dependence of the convergence rate on the clipping threshold C; the current statement appears to treat C as a fixed hyper-parameter without showing how it trades off against the privacy noise scale in the utility bound.

minor comments (2)

[System model] Notation for the wireless channel model (e.g., definition of sigma_w^2) should be introduced earlier and used consistently when stating the privacy-utility trade-off.
[Numerical results] The numerical experiments section should specify whether the reported privacy loss curves are obtained from the closed-form bound or from empirical estimation on the actual training trajectories.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below with clarifications from our analysis and indicate where revisions will be made to improve explicitness.

read point-by-point responses

Referee: [Privacy analysis section] Privacy analysis (around the composition theorem and Theorem on privacy loss): the headline claim that privacy loss converges to a constant rather than diverging with T rests on folding independent device selection probability p and additive wireless channel noise variance sigma_w^2 into the per-round leakage. The derivation should explicitly bound the sensitivity under these models and state the precise conditions (e.g., independence of selections across rounds) under which the constant bound holds; otherwise standard composition recovers linear or sqrt(T) growth.

Authors: Our privacy analysis incorporates device selection probability p and wireless noise variance sigma_w^2 directly into the per-round sensitivity and noise scale of the Gaussian mechanism. The effective sensitivity is p times the clipped gradient norm, and sigma_w^2 augments the total noise variance, yielding a per-round privacy parameter that is strictly smaller than the standard DP-FL case. Under independent selections across rounds (explicitly assumed in our model of random participation), the infinite-horizon composition is bounded by a geometric series summing to a finite constant. We agree the presentation can be tightened; we will add an explicit lemma stating the sensitivity bound and the independence condition in the revised privacy section. revision: yes
Referee: [Convergence guarantees section] Convergence analysis (section on non-convex rates with clipping): the guarantees for smooth non-convex objectives with gradient clipping require an explicit dependence of the convergence rate on the clipping threshold C; the current statement appears to treat C as a fixed hyper-parameter without showing how it trades off against the privacy noise scale in the utility bound.

Authors: The non-convex convergence theorem already expresses the rate in terms of C through both the clipping bias term (bounded by 2C) and the privacy noise variance, which is calibrated to the sensitivity that depends on C. The privacy-utility trade-off corollary then substitutes this noise scale into the final error bound. To make the dependence on C and its interaction with the privacy noise fully transparent, we will revise the theorem statement and add a short discussion paragraph explicitly showing the trade-off. revision: yes

Circularity Check

0 steps flagged

No significant circularity in privacy-convergence derivation

full rationale

The paper derives the constant privacy-loss bound by folding independent device selection probability p and wireless channel noise variance into the per-round privacy accounting, then applying standard advanced composition. This is a direct consequence of the stated system model rather than a self-referential definition or fitted parameter renamed as prediction. Convergence guarantees for smooth non-convex objectives with gradient clipping follow from conventional smoothness and Lipschitz assumptions without reducing to the inputs by construction. No load-bearing self-citations or uniqueness theorems imported from prior author work are required for the central claims.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Insufficient detail in abstract to enumerate all free parameters or invented entities; analysis rests on standard domain assumptions for smoothness and non-convexity in optimization plus wireless federated learning models.

axioms (2)

domain assumption General smooth non-convex loss objectives
Invoked to extend beyond restrictive convexity assumptions in prior work.
domain assumption Wireless channel and device selection models
Used to incorporate device selection and mini-batch sampling into privacy and convergence analysis.

pith-pipeline@v0.9.0 · 5420 in / 1351 out tokens · 35521 ms · 2026-05-15T08:08:04.713918+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

ε=2α p q c²/σ² min(∑(γ(t))², Φ) with Φ involving bounded diameter D; privacy loss converges to constant rather than diverging with T

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 1 internal anchor

[1]

With the growing prevalence of wireless networks, FL has been increasingly deployed over wireless channels, giving rise to wireless federated learning (WFL)

INTRODUCTION Federated learning (FL) has emerged as a standard paradigm for distributed learning, enabling multiple devices to collab- oratively train a shared model without exchanging raw data [1]. With the growing prevalence of wireless networks, FL has been increasingly deployed over wireless channels, giving rise to wireless federated learning (WFL). ...

work page
[2]

has become the de facto standard for ensuring individual privacy, offering strict theoretical guarantees. The key idea of DP is to inject carefully calibrated randomness, typically in the form of additive noise, to mask the contribution of any This work was supported by the Fundamental Research Funds for the Central Universities under Grants No. 207202520...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

System Model Learning Protocol.We consider an FL system that consists of a single-antenna server andnclients

PRELIMINARIES 2.1. System Model Learning Protocol.We consider an FL system that consists of a single-antenna server andnclients. Each deviceihas its own datasetD i, which is independent and identically dis- tributed (i.i.d.) across clients, and we assume that each device has the same number of samples. The server and clients co- operatively learn a global...

work page
[4]

Privacy Analysis In this subsection, we quantify the RDP loss for DPWFL

MAIN RESULTS 3.1. Privacy Analysis In this subsection, we quantify the RDP loss for DPWFL. Our theoretical result builds upon the recent work [14], which es- tablishes a rigorous privacy characterization of centralized learning under ideal communication settings. In DPWFL, the combination of inherent channel noise, mini-batch sampling, and device active s...

work page
[5]

NUMERICAL RESULTS In this section, we numerically evaluate the privacy levelϵ of the DPWFL withδ= 10 −5, and compare it against DP- NOMA [5], a representative baseline. Unless otherwise spec- ified, we set the device sampling ratep= 1, mini-batch sam- pling rateq= 1, clipping normc= 2, domain diameter D= 0.5, smoothness constantL= 1, channel noise scale σ...

work page
[6]

We showed that by lever- aging inherent channel noise and incorporating both device and mini-batch sampling, the privacy loss can converge to a constant

CONCLUSION In this paper, we provided a comprehensive analysis of pri- vacy and convergence for DPWFL. We showed that by lever- aging inherent channel noise and incorporating both device and mini-batch sampling, the privacy loss can converge to a constant. We further established the convergence result with gradient clipping, derived an explicit privacy–ut...

work page
[7]

Communication-efficient learning of deep networks from decentralized data,

Brendan McMahan, Eider Moore, Daniel Ram- age, Seth Hampson, and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. PMLR, 2017, pp. 1273–1282

work page 2017
[8]

Federated learning via over-the-air computation,

Kai Yang, Tao Jiang, Yuanming Shi, and Zhi Ding, “Federated learning via over-the-air computation,” IEEE transactions on wireless communications, vol. 19, no. 3, pp. 2022–2035, 2020

work page 2022
[9]

Comprehensive privacy analysis of deep learning: Pas- sive and active white-box inference attacks against cen- tralized and federated learning,

Milad Nasr, Reza Shokri, and Amir Houmansadr, “Comprehensive privacy analysis of deep learning: Pas- sive and active white-box inference attacks against cen- tralized and federated learning,” in2019 IEEE sympo- sium on security and privacy (SP). IEEE, 2019, pp. 739– 753

work page 2019
[10]

The algorithmic foundations of differential privacy,

Cynthia Dwork, Aaron Roth, et al., “The algorithmic foundations of differential privacy,”Foundations and Trends® in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014

work page 2014
[11]

Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,

Dongzhu Liu and Osvaldo Simeone, “Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 170–185, 2020

work page 2020
[12]

Private wireless federated learning with anonymous over-the-air compu- tation,

Burak Hasırcıo˘glu and Deniz G¨und¨uz, “Private wireless federated learning with anonymous over-the-air compu- tation,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5195–5199

work page 2021
[13]

On the differential pri- vacy in federated learning based on over-the-air com- putation,

Sangjun Park and Wan Choi, “On the differential pri- vacy in federated learning based on over-the-air com- putation,”IEEE Transactions on Wireless Communica- tions, vol. 23, no. 5, pp. 4269–4283, 2023

work page 2023
[14]

Dif- ferentially private over-the-air federated learning over mimo fading channels,

Hang Liu, Jia Yan, and Ying-Jun Angela Zhang, “Dif- ferentially private over-the-air federated learning over mimo fading channels,”IEEE Transactions on Wireless Communications, vol. 23, no. 8, pp. 8232–8247, 2024

work page 2024
[15]

Deep learning with differential privacy,

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang, “Deep learning with differential privacy,” inProceed- ings of the 2016 ACM SIGSAC conference on computer and communications security, 2016, pp. 308–318

work page 2016
[16]

Privacy amplification by itera- tion,

Vitaly Feldman, Ilya Mironov, Kunal Talwar, and Abhradeep Thakurta, “Privacy amplification by itera- tion,” in2018 IEEE 59th Annual Symposium on Foun- dations of Computer Science (FOCS). IEEE, 2018, pp. 521–532

work page 2018
[17]

Dif- ferential privacy dynamics of langevin diffusion and noisy gradient descent,

Rishav Chourasia, Jiayuan Ye, and Reza Shokri, “Dif- ferential privacy dynamics of langevin diffusion and noisy gradient descent,”Advances in Neural Infor- mation Processing Systems, vol. 34, pp. 14771–14781, 2021

work page 2021
[18]

Privacy of noisy stochastic gradient descent: More iterations without more privacy loss,

Jason Altschuler and Kunal Talwar, “Privacy of noisy stochastic gradient descent: More iterations without more privacy loss,”Advances in Neural Information Processing Systems, vol. 35, pp. 3788–3800, 2022

work page 2022
[19]

Wire- less federated learning with local differential privacy,

Mohamed Seif, Ravi Tandon, and Ming Li, “Wire- less federated learning with local differential privacy,” in2020 IEEE International Symposium on Information Theory (ISIT). IEEE, 2020, pp. 2604–2609

work page 2020
[20]

An improved privacy and utility anal- ysis of differentially private sgd with bounded domain and smooth losses,

Hao Liang, Wanrong Zhang, Xinlei He, Kaishun Wu, and Hong Xing, “An improved privacy and utility anal- ysis of differentially private sgd with bounded domain and smooth losses,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, 2026, vol. 40

work page 2026
[21]

R ´enyi differential privacy,

Ilya Mironov, “R ´enyi differential privacy,” in2017 IEEE 30th computer security foundations symposium (CSF). IEEE, 2017, pp. 263–275

work page 2017
[22]

On measures of entropy and informa- tion,

Alfr ´ed R ´enyi, “On measures of entropy and informa- tion,” inProceedings of the fourth Berkeley sympo- sium on mathematical statistics and probability, volume 1: contributions to the theory of statistics. University of California Press, 1961, vol. 4, pp. 547–562

work page 1961
[23]

Private selection from private candidates,

Jingcheng Liu and Kunal Talwar, “Private selection from private candidates,” inProceedings of the 51st Annual ACM SIGACT Symposium on Theory of Com- puting, 2019, pp. 298–309

work page 2019
[24]

Differential privacy as a perk: Federated learning over multiple-access fading channels with a multi-antenna base station,

Hao Liang, Haifeng Wen, Kaishun Wu, and Hong Xing, “Differential privacy as a perk: Federated learning over multiple-access fading channels with a multi-antenna base station,”arXiv preprint arXiv:2510.23463, 2025

work page arXiv 2025
[25]

Device schedul- ing for secure aggregation in wireless federated learn- ing,

Na Yan, Kezhi Wang, Kangda Zhi, Cunhua Pan, Kok Keong Chai, and H Vincent Poor, “Device schedul- ing for secure aggregation in wireless federated learn- ing,”IEEE Internet of Things Journal, vol. 11, no. 17, pp. 28851–28862, 2024

work page 2024
[26]

An improved analysis of per-sample and per-update clipping in federated learning,

Bo Li, Xiaowen Jiang, Mikkel N Schmidt, Tommy Sonne Alstrøm, and Sebastian U Stich, “An improved analysis of per-sample and per-update clipping in federated learning,” inThe International Conference on Learning Representations, 2024

work page 2024
[27]

On the privacy leakage of over-the-air federated learning over mimo fading channels,

Hang Liu, Jia Yan, and Ying-Jun Angela Zhang, “On the privacy leakage of over-the-air federated learning over mimo fading channels,” inGLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023, pp. 5274–5279

work page 2023

[1] [1]

With the growing prevalence of wireless networks, FL has been increasingly deployed over wireless channels, giving rise to wireless federated learning (WFL)

INTRODUCTION Federated learning (FL) has emerged as a standard paradigm for distributed learning, enabling multiple devices to collab- oratively train a shared model without exchanging raw data [1]. With the growing prevalence of wireless networks, FL has been increasingly deployed over wireless channels, giving rise to wireless federated learning (WFL). ...

work page

[2] [2]

has become the de facto standard for ensuring individual privacy, offering strict theoretical guarantees. The key idea of DP is to inject carefully calibrated randomness, typically in the form of additive noise, to mask the contribution of any This work was supported by the Fundamental Research Funds for the Central Universities under Grants No. 207202520...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[3] [3]

System Model Learning Protocol.We consider an FL system that consists of a single-antenna server andnclients

PRELIMINARIES 2.1. System Model Learning Protocol.We consider an FL system that consists of a single-antenna server andnclients. Each deviceihas its own datasetD i, which is independent and identically dis- tributed (i.i.d.) across clients, and we assume that each device has the same number of samples. The server and clients co- operatively learn a global...

work page

[4] [4]

Privacy Analysis In this subsection, we quantify the RDP loss for DPWFL

MAIN RESULTS 3.1. Privacy Analysis In this subsection, we quantify the RDP loss for DPWFL. Our theoretical result builds upon the recent work [14], which es- tablishes a rigorous privacy characterization of centralized learning under ideal communication settings. In DPWFL, the combination of inherent channel noise, mini-batch sampling, and device active s...

work page

[5] [5]

NUMERICAL RESULTS In this section, we numerically evaluate the privacy levelϵ of the DPWFL withδ= 10 −5, and compare it against DP- NOMA [5], a representative baseline. Unless otherwise spec- ified, we set the device sampling ratep= 1, mini-batch sam- pling rateq= 1, clipping normc= 2, domain diameter D= 0.5, smoothness constantL= 1, channel noise scale σ...

work page

[6] [6]

We showed that by lever- aging inherent channel noise and incorporating both device and mini-batch sampling, the privacy loss can converge to a constant

CONCLUSION In this paper, we provided a comprehensive analysis of pri- vacy and convergence for DPWFL. We showed that by lever- aging inherent channel noise and incorporating both device and mini-batch sampling, the privacy loss can converge to a constant. We further established the convergence result with gradient clipping, derived an explicit privacy–ut...

work page

[7] [7]

Communication-efficient learning of deep networks from decentralized data,

Brendan McMahan, Eider Moore, Daniel Ram- age, Seth Hampson, and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. PMLR, 2017, pp. 1273–1282

work page 2017

[8] [8]

Federated learning via over-the-air computation,

Kai Yang, Tao Jiang, Yuanming Shi, and Zhi Ding, “Federated learning via over-the-air computation,” IEEE transactions on wireless communications, vol. 19, no. 3, pp. 2022–2035, 2020

work page 2022

[9] [9]

Comprehensive privacy analysis of deep learning: Pas- sive and active white-box inference attacks against cen- tralized and federated learning,

Milad Nasr, Reza Shokri, and Amir Houmansadr, “Comprehensive privacy analysis of deep learning: Pas- sive and active white-box inference attacks against cen- tralized and federated learning,” in2019 IEEE sympo- sium on security and privacy (SP). IEEE, 2019, pp. 739– 753

work page 2019

[10] [10]

The algorithmic foundations of differential privacy,

Cynthia Dwork, Aaron Roth, et al., “The algorithmic foundations of differential privacy,”Foundations and Trends® in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014

work page 2014

[11] [11]

Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,

Dongzhu Liu and Osvaldo Simeone, “Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 170–185, 2020

work page 2020

[12] [12]

Private wireless federated learning with anonymous over-the-air compu- tation,

Burak Hasırcıo˘glu and Deniz G¨und¨uz, “Private wireless federated learning with anonymous over-the-air compu- tation,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5195–5199

work page 2021

[13] [13]

On the differential pri- vacy in federated learning based on over-the-air com- putation,

Sangjun Park and Wan Choi, “On the differential pri- vacy in federated learning based on over-the-air com- putation,”IEEE Transactions on Wireless Communica- tions, vol. 23, no. 5, pp. 4269–4283, 2023

work page 2023

[14] [14]

Dif- ferentially private over-the-air federated learning over mimo fading channels,

Hang Liu, Jia Yan, and Ying-Jun Angela Zhang, “Dif- ferentially private over-the-air federated learning over mimo fading channels,”IEEE Transactions on Wireless Communications, vol. 23, no. 8, pp. 8232–8247, 2024

work page 2024

[15] [15]

Deep learning with differential privacy,

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang, “Deep learning with differential privacy,” inProceed- ings of the 2016 ACM SIGSAC conference on computer and communications security, 2016, pp. 308–318

work page 2016

[16] [16]

Privacy amplification by itera- tion,

Vitaly Feldman, Ilya Mironov, Kunal Talwar, and Abhradeep Thakurta, “Privacy amplification by itera- tion,” in2018 IEEE 59th Annual Symposium on Foun- dations of Computer Science (FOCS). IEEE, 2018, pp. 521–532

work page 2018

[17] [17]

Dif- ferential privacy dynamics of langevin diffusion and noisy gradient descent,

Rishav Chourasia, Jiayuan Ye, and Reza Shokri, “Dif- ferential privacy dynamics of langevin diffusion and noisy gradient descent,”Advances in Neural Infor- mation Processing Systems, vol. 34, pp. 14771–14781, 2021

work page 2021

[18] [18]

Privacy of noisy stochastic gradient descent: More iterations without more privacy loss,

Jason Altschuler and Kunal Talwar, “Privacy of noisy stochastic gradient descent: More iterations without more privacy loss,”Advances in Neural Information Processing Systems, vol. 35, pp. 3788–3800, 2022

work page 2022

[19] [19]

Wire- less federated learning with local differential privacy,

Mohamed Seif, Ravi Tandon, and Ming Li, “Wire- less federated learning with local differential privacy,” in2020 IEEE International Symposium on Information Theory (ISIT). IEEE, 2020, pp. 2604–2609

work page 2020

[20] [20]

An improved privacy and utility anal- ysis of differentially private sgd with bounded domain and smooth losses,

Hao Liang, Wanrong Zhang, Xinlei He, Kaishun Wu, and Hong Xing, “An improved privacy and utility anal- ysis of differentially private sgd with bounded domain and smooth losses,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, 2026, vol. 40

work page 2026

[21] [21]

R ´enyi differential privacy,

Ilya Mironov, “R ´enyi differential privacy,” in2017 IEEE 30th computer security foundations symposium (CSF). IEEE, 2017, pp. 263–275

work page 2017

[22] [22]

On measures of entropy and informa- tion,

Alfr ´ed R ´enyi, “On measures of entropy and informa- tion,” inProceedings of the fourth Berkeley sympo- sium on mathematical statistics and probability, volume 1: contributions to the theory of statistics. University of California Press, 1961, vol. 4, pp. 547–562

work page 1961

[23] [23]

Private selection from private candidates,

Jingcheng Liu and Kunal Talwar, “Private selection from private candidates,” inProceedings of the 51st Annual ACM SIGACT Symposium on Theory of Com- puting, 2019, pp. 298–309

work page 2019

[24] [24]

Differential privacy as a perk: Federated learning over multiple-access fading channels with a multi-antenna base station,

Hao Liang, Haifeng Wen, Kaishun Wu, and Hong Xing, “Differential privacy as a perk: Federated learning over multiple-access fading channels with a multi-antenna base station,”arXiv preprint arXiv:2510.23463, 2025

work page arXiv 2025

[25] [25]

Device schedul- ing for secure aggregation in wireless federated learn- ing,

Na Yan, Kezhi Wang, Kangda Zhi, Cunhua Pan, Kok Keong Chai, and H Vincent Poor, “Device schedul- ing for secure aggregation in wireless federated learn- ing,”IEEE Internet of Things Journal, vol. 11, no. 17, pp. 28851–28862, 2024

work page 2024

[26] [26]

An improved analysis of per-sample and per-update clipping in federated learning,

Bo Li, Xiaowen Jiang, Mikkel N Schmidt, Tommy Sonne Alstrøm, and Sebastian U Stich, “An improved analysis of per-sample and per-update clipping in federated learning,” inThe International Conference on Learning Representations, 2024

work page 2024

[27] [27]

On the privacy leakage of over-the-air federated learning over mimo fading channels,

Hang Liu, Jia Yan, and Ying-Jun Angela Zhang, “On the privacy leakage of over-the-air federated learning over mimo fading channels,” inGLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023, pp. 5274–5279

work page 2023