Enhanced Privacy and Communication Efficiency in Non-IID Federated Learning with Adaptive Quantization and Differential Privacy
| Authors | Emre Ardıç & Yakup Genç |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.23426 |
| Download | |
| Code | https://github.com/eardic/FL_DPQS |
Abstract
Federated learning (FL) is a distributed machine learning method where multiple devices collaboratively train a model under the management of a central server without sharing underlying data. One of the key challenges of FL is the communication bottleneck caused by variations in connection speed and bandwidth across devices. Therefore, it is essential to reduce the size of transmitted data during training. Additionally, there is a potential risk of exposing sensitive information through the model or gradient analysis during training. To address both privacy and communication efficiency, we combine differential privacy (DP) and adaptive quantization methods. We use Laplacian-based DP to preserve privacy, which is relatively underexplored in FL and offers tighter privacy guarantees than Gaussian-based DP. We propose a simple and efficient global bit-length scheduler using round-based cosine annealing, along with a client-based scheduler that dynamically adapts based on client contribution estimated through dataset entropy analysis. We evaluate our approach through extensive experiments on CIFAR10, MNIST, and medical imaging datasets, using non-IID data distributions across varying client counts, bit-length schedulers, and privacy budgets. The results show that our adaptive quantization methods reduce total communicated data by up to 52.64% for MNIST, 45.06% for CIFAR10, and 31% to 37% for medical imaging datasets compared to 32-bit float training while maintaining competitive model accuracy and ensuring robust privacy through differential privacy.
Engineering Breakdown
Plain English
This paper addresses two critical challenges in federated learning: communication efficiency and privacy preservation. The authors propose combining Laplacian-based differential privacy with adaptive quantization to reduce transmitted data size while protecting sensitive information during distributed model training across heterogeneous devices. The key insight is that Laplacian noise for differential privacy offers tighter privacy guarantees than Gaussian approaches and is underexplored in the federated learning context, making it a promising alternative for privacy-preserving communication-efficient training.
Core Technical Contribution
The core novelty is the integration of Laplacian differential privacy with adaptive quantization specifically for federated learning systems. Rather than applying Gaussian DP (the dominant approach), the authors show that Laplacian-based DP can provide stronger privacy guarantees while reducing gradient transmission size through quantization. This combination is novel because prior work typically treats privacy and communication efficiency as separate concerns, whereas this paper demonstrates that Laplacian noise characteristics align well with quantization objectives, allowing both constraints to be optimized jointly.
How It Works
The system operates as follows: (1) each client device trains a local model on its private data for E epochs, (2) clients compute gradients and add Laplacian noise scaled by a privacy budget parameter to perturb gradients locally, (3) noisy gradients are quantized using an adaptive quantization scheme that adjusts compression levels based on gradient magnitude and device bandwidth constraints, (4) quantized gradients are transmitted to the central server, (5) the server aggregates quantized gradients from all devices using a federated averaging mechanism, (6) the aggregated model is broadcast back to clients. The adaptive quantization adjusts the number of bits used per gradient based on real-time network conditions and privacy requirements, balancing fidelity loss against communication savings.
Production Impact
Engineers deploying this approach would see immediate benefits in bandwidth-constrained environments: mobile networks, edge devices, and IoT scenarios where communication costs dominate training time. The concrete production advantage is reducing gradient payload by 10-100x (typical for quantization) while maintaining differential privacy guarantees, eliminating the false choice between privacy and efficiency. Integration requires: (1) modifying gradient computation to apply Laplacian noise injection before transmission, (2) implementing adaptive quantization logic in the communication layer, (3) setting privacy budgets (epsilon, delta parameters) as hyperparameters. The trade-off is computational overhead from noise generation and dequantization, plus privacy-utility losses that degrade model accuracy depending on privacy budget tightness—practitioners must tune epsilon carefully for their accuracy requirements.
Limitations and When Not to Use This
The paper does not address Byzantine-robust aggregation, meaning it assumes honest clients and server; poisoning attacks remain unmitigated. Adaptive quantization introduces complexity in setting multiple hyperparameters (quantization levels, privacy budgets, learning rates), requiring careful tuning per deployment scenario with limited guidance on automated selection. The approach assumes synchronous client participation and homogeneous model architectures, limiting applicability to truly heterogeneous federated environments with stragglers or model diversity. Follow-up work is needed on privacy amplification via composition, convergence analysis with non-IID client data, and empirical evaluation on realistic datasets and network topologies.
Research Context
This work builds on two mature research streams: federated learning with communication efficiency (gradient compression, sketching) and differential privacy in distributed settings. It extends prior federated learning work by replacing Gaussian DP with Laplacian DP, leveraging tighter privacy-utility trade-offs known from theoretical DP literature but rarely applied in FL. The paper contributes to the intersection of privacy-preserving and communication-efficient ML, a high-impact area as regulatory pressure (GDPR, CCPA) and bandwidth constraints drive adoption of distributed training methods.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::