Abstract
Cell-free massive MIMO (CF mMIMO) is a promising next-generation wireless architecture to realize federated learning (FL). However, sensitive information of user equipments (UEs) may be exposed to the involved access points or the central processing unit in practice. To guarantee data privacy, effective privacy-preserving mechanisms are defined in this paper. In particular, we demonstrate and characterize the possibility of exploiting the inherent quantization error, caused by low-resolution analog-to-digital converters (ADCs) and digital-to-analog converters (DACs), for privacy-preserving in a FL CF mMIMO system. Furthermore, to reduce the required uplink training time in such a system, a stochastic non-convex design problem that jointly optimizes the transmit power and the data rate is formulated. To address the problem at hand, we propose a novel power control method by utilizing the successive convex approximation (SCA) approach to obtain a suboptimal solution. Besides, an asynchronous protocol is established for mitigating the straggler effect to facilitate FL. Numerical results show that compared with the conventional full power transmission, adopting the proposed power control method can effectively reduce the uplink training time under various practical system settings. Also, our results unveil that our proposed asynchronous approach can reduce the waiting time at the central processing unit for receiving all user information, as there are no stragglers that require a long time to report their local updates.
Introduction
Federated Learning (FL) has emerged as a cornerstone for privacy-preserving AI, allowing local devices to train models without sharing raw data. However, the wireless transmission of model gradients remains a vulnerability, susceptible to "model inversion attacks" where adversaries reconstruct private data from updates. Moreover, practical FL deployments face significant hurdles in communication latency and the "straggler effect," where the slowest device bottlenecks the entire network's progress.
This paper proposes a framework that integrates FL with Cell-Free Massive MIMO (CF mMIMO). In CF mMIMO, a large number of distributed access points (APs) cooperatively serve users, providing the macro-diversity necessary for reliable high-speed connections. By leveraging this architecture, the authors introduce a novel way to achieve "privacy for free" while optimizing network efficiency.
System Architecture
The proposed system operates through a coordinated cycle between a Central Processing Unit (CPU), multiple Access Points (APs), and User Equipments (UEs). The workflow is divided into four distinct steps that manage the flow of model parameters and local updates.
As shown in the architecture above, the process begins with Step 1, where the CPU sends the global model to UEs via downlink communication. In Step 2, UEs perform local computation to update the model based on their private data. Step 3 involves the UEs uploading these local updates back to the CPU through the APs. Finally, in Step 4, the CPU performs global computation to aggregate the updates and refine the global model. The wireless links between UEs and APs, and the fronthaul between APs and the CPU, are the primary focus for optimization and privacy protection.
Methods
Quantization-Based Differential Privacy (DP)
Traditionally, Differential Privacy is achieved by adding artificial Gaussian noise to gradients, which consumes power and can degrade model accuracy. This paper flips the script by utilizing the inherent quantization noise generated by low-resolution hardware—specifically, ADCs at the APs and DACs at the UEs. The authors derive a closed-form $(\epsilon, \delta)$-DP condition, proving that this unavoidable hardware imperfection can serve as a formal privacy guarantee, reducing the need for additional noise injection.
Joint Power and Rate Optimization
To minimize the total uplink training time, the authors formulate a stochastic non-convex optimization problem. Given the complexity of the wireless environment, they employ Successive Convex Approximation (SCA). This method transforms the difficult non-convex problem into a series of simpler convex sub-problems, resulting in a power control strategy that significantly outperforms standard full-power transmission.
Asynchronous FL Protocol
To eliminate the straggler effect, an asynchronous protocol is introduced. Unlike synchronous FL, the CPU does not wait for every user to finish. By implementing "lag tolerance" and "lag percent" parameters, the system can proceed using updates from the fastest users. This ensures that the global model continues to evolve without being held hostage by users with poor link quality, while still allowing stragglers to sync up eventually to prevent model poisoning.
Key Results and Insights
- Hardware Imperfections as a Feature: The study reveals a "sweet spot" for hardware resolution. Using 3-bit ADCs performs nearly as well as high-resolution 10-bit hardware, suggesting that low-cost, low-power hardware is sufficient for efficient, private FL.
- Significant Speedups: The proposed power control method reduces uplink training time by 35% with 1-bit quantization and 21% with 10-bit hardware. When combined with the asynchronous protocol, the system was found to be 2.5 times faster than standard synchronous modes.
- Macro-Diversity Gains: Increasing the number of APs consistently reduced training time. The increased macro-diversity of the CF mMIMO architecture helps smooth out interference, making the overall FL process more robust.
Conclusion
By repurposing hardware quantization noise for privacy and implementing an asynchronous communication protocol, this research provides a viable path for deploying FL in next-generation wireless networks. Future work will look to relax assumptions regarding perfect Channel State Information (CSI) and explore the impact of constrained fronthaul links on the privacy-utility tradeoff.
Summary updated