Privacy-Preserving Personalized Federated Learning Against Poisoning and Inference Attacks in Intelligent Enterprises

Abstract

Data-driven decision-making is a key enabler of intelligent enterprises, yet centralized machine learning (ML) poses privacy and efficiency risks. Federated learning (FL) offers a decentralized solution. Still, it faces 1) high communication cost, 2) vulnerability to poisoning and inference attacks, and 3) degraded performance on non-independent and identically distributed (non-IID) data. This paper presents FedAnil+++, a privacy-preserving and communication-efficient FL model with three phases: (i) dynamic gradient compression, (ii) multi-layer protection via cosine similarity (CS) filtering, CKKS-FHE, and blockchain aggregation, and (iii) personalized FL (PFL) through clustering and adversarial learning. Experiments on FEMNIST, CIFAR-10, and TON-IoT show that FedAnil+++ boosts accuracy by 15.75%, cuts communication by 25.87%, and reduces computation by 4.5% over baselines. Privacy metrics, Kullback–Leibler (KL) divergence, Shannon entropy, and correlation analysis confirm that CKKS-FHE effectively conceals sensitive patterns. The implementation of FedAnil+++ is publicly available on GitHub https://github.com/rezafotohi/FedAnilPlusPlusPlus