Scalable and Private Federated Learning Using Distributed Differential Privacy and Secure Aggregation

Wenjing Wei; Farid Nait-Abdesselam; Alla Jammine

cs.CRApr 8, 2026

Scalable and Private Federated Learning Using Distributed Differential Privacy and Secure Aggregation

Authors: Wenjing Wei, Farid Nait-Abdesselam, Alla Jammine

Organizations: Université Paris Cité, Paris, France

Abstract

This article presents DDP-SA, a scalable privacy-preserving federated learning framework that jointly leverages client-side local differential privacy (LDP) and full-threshold additive secret sharing (ASS) for secure aggregation. Unlike existing methods that rely solely on differential privacy or on secure multi-party computation (MPC), DDP-SA integrates both techniques to deliver stronger end-to-end privacy guarantees while remaining computationally practical. The framework introduces a two-stage protection mechanism: clients first perturb their local gradients with calibrated Laplace noise, then decompose the noisy gradients into additive secret shares that are distributed across multiple intermediate servers. This design ensures that (i) no single compromised server or communication channel can reveal any information about individual client updates, and (ii) the parameter server reconstructs only the aggregated noisy gradient, never any client-specific contribution. Extensive experiments show that DDP-SA achieves substantially higher model accuracy than standalone LDP while providing stronger privacy protection than MPC-only approaches. The proposed framework scales linearly with the number of participants and offers a practical, privacy-preserving solution for federated learning applications with controllable computational and communication overhead.

Figures & tables

Federated Scenario	Publications	Year	DP Model	Neighborhood Level	Perturbation Mechanism	$\textbf{CM}^{1}$	Downstream Tasks	$\begin{tabular}[c]{@{}c@{}}Model\\ Architecture\end{tabular}^{2}$	Clients Number	$\epsilon$	$\delta$
	Chen et al. [ 27 ]	2024			Gaussian	tCDP	Classification	LR, Shallow CNN	100	0.3	$10^{-2}$
	malekmohammadi et al. [ 28 ]	2024			Gaussian	AC	Classification	CNN	[20,60]	[0.5,5]	$10^{-4}$
	Liu et al. [ 29 ]	2024			Gaussian	RDP	Classification	CNN	10	[0.1,10]	$10^{-3}$
	Ling et al. [ 30 ]	2024			Gaussian	RDP	Classification	Shallow CNN	10	[1.5,5.5]	$10^{-5}$
	Xiang et al. [ 34 ]	2023			Gaussian	MA	Classification	Shallow CNN,LSTM	[10,20]	[0.12,2]	$[10^{-2},10^{-5}]$
	Ruan et al. [ 31 ]	2023			Gaussian	RDP	Classification	Shallow CNN, LSTM	[3,10]	[0.25,2]	$[10^{-4},10^{-5}]$

TABLE I: An Overview Study of Differentially Private FL [ 33 ]

Fig. 1: DDP-SA Framework diagram.

Fig. 2: DDP-SA workflow, general scalable framework with $n$ clients, $m$ intermediate servers, and $d$ -dimensional parameters.

Component	Operation	Time (ms)	Percentage of Total	Scalability
LDP	Gradient Clipping	547.95	92.07%	$O(d)$
	Noise Generation	0.24	0.04%	$O(d)$
	Noise Addition	0.05	0.01%	$O(d)$
MPC	Fixed-Point Encoding	0.22	0.04%	$O(d)$
	Secret Sharing	0.53	0.09%	$O(d\cdot m)$
	Share Transmission	46.13	7.75%	$O(d\cdot m)$

TABLE II: Computational Overhead Breakdown per Client per Round

Component	Direction	Bytes	Percentage of Total	Scalability
LDP	PS to Client	$4d$	50.0%	$O(d)$
LDP	Client to PS	$4d$	50.0%	$O(d)$
MPC	PS to Client	$4d$	25.0%	$O(d)$
	Client to Intermediate Servers (excluded)	$4d\cdot m$	-	$O(d\cdot m)$
	Intermediate Servers to PS	$4d\cdot m$	75.0% (for $m=3$ )	$O(d\cdot m)$
DDP-SA	PS to Client	$4d$	25.0%	$O(d)$

TABLE III: Communication Overhead Breakdown per Client per Round

Fig. 3: Number of communication rounds for different defensive mechanisms.

Fig. 4: Number of parameters uploaded per client for different defensive mechanisms. Results shown for

m=3

Fig. 5: Total time to convergence for different defensive mechanisms.

Fig. 6: Average training time per round for different defensive mechanisms.

Panel of Figure 10 — Fig. 7: Accuracy for different defensive mechanisms. (a) Test loss. (b) Test $\text{R}^{2}$ .

Panel of Figure 11 — Fig. 8: Accuracy as a function of privacy budget $\epsilon$ . (a) Test loss vs. $\epsilon$ . (b) Test $\text{R}^{2}$ vs. $\epsilon$ .

Panel of Figure 12 — Fig. 9: Accuracy as a function of the number of clients $n$ . (a) Test loss vs. $n$ . (b) Test $\text{R}^{2}$ vs. $n$ .

Fig. 10: Training loss as a function of the number of communication rounds

T

Privacy-Preserving Federated Learning

See all 41 papers

Privacy-enhanced federated learning via asynchronous aggregation and local differential perturbation
Sep 14, 2026Zhen Zhong, Shini Yang, Liesheng WeiFederated Learning Aggregation Asynchronous Federated Learning
Georgetown University, Washington, D.C., USA · LinkedIn, CA, USA · Shanghai Ocean University, Shanghai, China
DisAgg: Distributed Aggregators for Efficient Secure Aggregation in Federated Learning
May 13, 2026Haaris Mehmood, Giorgos Tatsis, Dimitrios Alexopoulos +4Federated Learning Aggregation Communication-Efficient Distributed Training
Samsung R&D Institute UK (SRUK) · Information Technologies Institute (CERTH-ITI)
An Adaptive Gradient Clipping and Noise Injection Mechanism for Differentially Private Federated Learning
Aug 15, 2026Wenjing Wei, Alla Jammine, Farid Nait-AbdesselamPrivacy-Preserving Federated Learning Differentially Private Federated Learning
Universit´e Paris Cit´e

Sep 14, 2026J/K move · Enter open · S save

Privacy-enhanced federated learning via asynchronous aggregation and local differential perturbation

Zhen Zhong, Shini Yang, Liesheng Wei

Georgetown University, Washington, D.C., USA · LinkedIn, CA, USA · Shanghai Ocean University, Shanghai, China

Federated Learning Aggregation Asynchronous Federated Learning Differentially Private Federated Learning Differential Privacy

Open paper AskPDF arXiv

This study proposes a privacy-enhanced federated learning framework to address secure collaborative training in distributed data environments. The framework integrates Dynamic Differential Privacy (DDP), lightweight Homomorphic Encryption (HE), and Local Differential Privacy (LDP) mechanisms to ensure data privacy protection during model training. Additionally, the framework employs an asynchronous aggregation strategy with version control to support distributed training in asynchronous environments. Experimental validation on the CIFAR-10 and Purchase-100 benchmark datasets demonstrates that the method maintains high classification accuracy (up to 82.6%) even under stringent privacy constraints (ε = 0.1), while reducing communication overhead by 21.3% compared to FedAvg. Experimental results demonstrate that this framework effectively balances privacy protection and model performance in distributed machine learning scenarios, providing a scalable technical foundation for large-scale distributed collaborative computing.

Abstract

Privacy-Preserving Federated Learning

Explore similar work

Privacy-enhanced federated learning via asynchronous aggregation and local differential perturbation

DisAgg: Distributed Aggregators for Efficient Secure Aggregation in Federated Learning

An Adaptive Gradient Clipping and Noise Injection Mechanism for Differentially Private Federated Learning