Security and Privacy Protection Methods for Federated Learning for Big Data

Abstract

In the big data environment, federated learning faces multiple challenges such as privacy leakage, poisoning attacks, and communication overload. Existing methods, mostly functioning as point defenses, struggle to simultaneously balance security, efficiency, and utility. This research aims to construct a multi-layered federated learning security system encompassing “source protection, process defense, and global optimization”. The study builds a cloud-edge-end collaborative architecture, integrating differential privacy with Shamir’s secret sharing to achieve data source perturbation and gradient share transmission. Through Mixup data augmentation combined with gradient clustering, it proactively detects poisoning attacks and introduces federated unlearning to remediate malicious impacts post factum. Based on static Bayesian games, it dynamically allocates privacy budgets to achieve a Nash equilibrium between personalized privacy and model utility. Experiments conducted on the CIFAR-10 and FEMNIST datasets, using a convolutional neural network as the base model and comparing it with the Vanilla FL model and module ablation versions, demonstrate the following: the FAA achieves a communication overhead of only 91.9 MB with 75 clients. Mixup combined with gradient clustering maintains an accuracy of 69.7% under 24% poisoning attacks. The game-theoretic framework attains a privacy–utility balance coefficient of up to 0.91. In complex dynamic scenarios, the multi-layered framework achieves an accuracy of 76.9%. This system exhibits robust security and adaptability under various attacks, providing a systematic solution for the practical deployment of federated learning.