Quantization in Federated Learning: Methods, Challenges and Future Directions

Research Context and Approach

Federated Learning (FL) enables collaborative model training across decentralized devices while preserving data privacy. However, the paradigm faces significant scalability hurdles, including high communication overhead, device heterogeneity, and the statistical challenges of non-IID (non-independent and identically distributed) data. Quantization—the process of reducing the numerical precision of model parameters or gradients—has emerged as a primary strategy to mitigate these bottlenecks.

This paper presents the first systematic literature review (SLR) focused exclusively on quantization within the FL ecosystem. Using a PRISMA-based methodology, the authors synthesize existing research to create a novel taxonomy. This taxonomy categorizes quantization methods based on training stages (e.g., post-training vs. quantization-aware training), the objects being quantized (parameters vs. gradients), encoding structures, and precision strategies.

Key Findings and Taxonomy

The authors argue that quantization is not merely a compression tool but a critical system-level design choice. Their taxonomy evaluates quantization across six dimensions: client heterogeneity, aggregation consistency, communication-scheduling, non-IID robustness, privacy/security integration, and hardware/energy co-optimization.

One notable focus is Post-Training Quantization (PTQ), specifically Dynamic Range Quantization (DRQ). DRQ is highlighted for its practicality in edge deployments, as it quantizes weights to lower bit-widths (e.g., INT8) while maintaining floating-point activations. This approach offers a favorable trade-off between communication reduction and computational overhead without requiring expensive retraining. However, the authors note that the effectiveness of these methods is highly sensitive to the representativeness of calibration data and the degree of statistical heterogeneity among clients.

By consolidating fragmented research, this survey provides a roadmap for practitioners and researchers. It identifies critical gaps, such as the lack of standardized benchmarks and the need for adaptive quantization strategies that can co-evolve with aggregation and scheduling policies. The paper serves as a foundational guide for designing robust, energy-efficient, and scalable FL systems suitable for real-world IoT and mobile environments.