A Comprehensive Survey of Privacy-preserving Federated Learning

Abstract

The past four years have witnessed the rapid development of federated learning (FL). However, new privacy concerns have also emerged during the aggregation of the distributed intermediate results. The emerging privacy-preserving FL (PPFL) has been heralded as a solution to generic privacy-preserving machine learning. However, the challenge of protecting data privacy while maintaining the data utility through machine learning still remains. In this article, we present a comprehensive and systematic survey on the PPFL based on our proposed 5W-scenario-based taxonomy. We analyze the privacy leakage risks in the FL from five aspects, summarize existing methods, and identify future research directions.

Applying Federated Learning in Software-Defined Networks: A Survey

“…FL approaches have been surveyed in many popular research scenarios, as listed in Table 1. In regard to security issues, Lyu et al [20] summarized the potential threats and defenses, and Yin et al [21] and Mothukuri et al [22] offered comprehensive surveys on data privacy and security. While Khan et al [23] provided a comprehensive survey on FL approaches, applications, and challenges over IoT, Pham et al presented an exhaustive review on FL over industrial IoT (IIoT) [24], Imteaj et al [25] surveyed the existing problems and solutions considering resource-constrained IoT devices.…”

A lightweight federated learning based privacy preserving B5G pandemic response network using unmanned aerial vehicles: A proof-of-concept

“…Categorization and security analysis of federated learning setups were elaborated in [20] while the privacy robustness and efficiency in the presence of an adversarial server was considered in [21] . An extensive investigation on privacy preserving federated learning was carried out in [22] that identified the potential privacy leakage risks in federated learning. However, most of these works are theoretical in nature and do not offer QoS (e.g., bandwidth and delay improvement) along with the implied privacy that are required for B5G networks as explored in the coauthor’s earlier work in [23] by leveraging the theoretical findings on customizing federated learning with asynchronous temporal features in [24] .…”

“… Reference Federated learning technique and scenario B5G-ready? [8] , [12] , [14] , [18] , [20] , [22] Horizontal, vertical, and hybrid data partitioning and/or cryptographic, perturbative, and anonymization techniques for privacy No [9] , [11] , [13] , [15] , [16] , [17] , [21] Specific focus on privacy-preservation in federated learning setups without any wireless communication parameter No [22] Privacy leakage scenarios with external and external attackers, active/passive attacks, inference attacks No [24] Temporally asynchronous weight update of federated learning No [19] , [23] Federated machine learning to address energy, bandwidth, delay and data privacy concerns in wireless communications by performing decentralized model training Yes …”

Fairness, Integrity, and Privacy in a Scalable Blockchain-based Federated Learning System

“…For example [53] developed a FL system for vehicular networks that combines LDP with gradient descent to avoid attacks that leak private information from publicly available ML model updates. Instead of perturbing model parameters, LDP noise can also be added to the training data [7]. However, this approach cannot provide privacy protection since it is not sufficient to make any single record unnoticeable [54].…”

“…Both traditional ML and distributed ML approaches assume that their training data is centralized by nature, preventing the applicability of ML approaches to domains in which data is sensitive and distributed at the same time. To avoid that ML approaches must rely on data to which only a centralized organization or individual has full access, federated machine learning (FL) can aggregate the less sensitive ML models that were independently and locally trained by individual clients [6,7]. Consequently, FL can enable the use of ML applications in domains with strong privacy requirements and contribute to solving the challenge of limited access to sensitive data without invading participating clients' privacy [8,9].…”