Privacy-Preserving Federated Learning Across Heterogeneous Distributed Datasets: Quantum-Inspired Mechanisms, Explainable Classifiers and Multi-Domain Empirical Evidence

Abstract

This paper addresses vulnerabilities in standard federated learning aggregation against gradient inversion attacks by evaluating three layered privacy-enhancing mechanisms: classical differential privacy (DP) noise injection, homomorphic encryption, and a novel quantum-inspired random unitary rotation of embedding vectors. Benchmarked across distributed threat-detection tasks specifically Bangla smishing classification and transaction card fraud detection the quantum-inspired rotation mechanism achieves the optimal privacy-utility trade-off. It preserves 97.3% of centralized model accuracy while introducing a minimal 4.1% per-round communication overhead, bypassing the steep accuracy penalties of classical DP and the high computational latency of homomorphic encryption. Furthermore, the paper introduces a federated SHAP protocol that permits regulatory-grade explainability auditing via masked attribution vectors without exposing local client data distributions. A hierarchical topology and drift-triggered re-aggregation schedule are also deployed to improve scalability and combat temporal concept drift.