Journal article · 2025
Privacy-Aware Compression for Federated Learning: A Comprehensive Survey of Mechanisms, Methods, and Trade-offs
International Journal of Latest Research in Engineering and Technology (IJLRET), vol. 11 (12), 7–21 ·
Abstract
Federated learning (FL) enables collaborative model training across distributed clients while preserving data privacy by avoiding raw data transmission. However, FL systems face critical challenges in balancing three competing objectives: privacy protection, communication efficiency, and model accuracy. This comprehensive survey reviews the state-of-the-art methods and mechanisms designed to address these challenges through privacy-aware compression. We systematically reviewed 50 seminal papers published between 2018–2025 across leading venues using a structured search methodology. The surveyed approaches are organized into four primary families: (i) numerical mechanism design approaches (e.g., Minimum Variance Unbiased and Interpolated MVU mechanisms) that jointly optimize noise allocation and bit budgets; (ii) compression-based methods (quantisation, sparsification, compressive sensing, and learned auto encoders) with privacy guarantees; (iii) incentive-compatible game-theoretic frameworks (e.g., auction and contract theory) for personalised privacy budgets; and (iv) real-world implementations on edge devices and IoT systems. Empirical evidence demonstrates that well-designed combinations of these approaches can achieve communication reductions of 50–90% while maintaining model accuracy within 1–3% of non-private baselines under formal privacy constraints (e.g., ε ≤ 2 differential privacy). We identify persistent challenges including non-linear privacy-utility-communication trade-offs, robustness to adaptive adversaries, handling client heterogeneity, and deployment feasibility. Finally, we discuss open research directions including Byzantine-robust compression, real-world regulatory compliance, and adaptive online compression schemes. This survey provides a unified perspective on how mechanism design, cryptographic, and learning-theoretic techniques can be synergistically combined to enable practical, provably private, and communication-efficient federated learning systems.
Nankai University, Tianjin, China.