Recently, Professor Cui Haixia's team from the School of Electronic Science and Engineering (School of Microelectronics), Faculty of Engineering, South China Normal University, has made significant progress in research on efficient and privacy-preserving model training for distributed edge devices. The related work, titled "Communication-Efficient Federated Learning for Edge Computing with Gradient Leakage Defense," has been published in IEEE Journal on Selected Areas in Communications (IEEE JSAC), a top-tier journal in the field of computer networking and communications (ranked in the Chinese Academy of Sciences Zone 1). The first author is Yang Xihong, a master's student enrolled in 2023, with Professor Cui Haixia as the corresponding author. South China Normal University is the primary affiliation.

    In recent years, federated learning has emerged as a privacy-preserving framework capable of training models across edge devices. However, in edge computing environments characterized by resource heterogeneity and unstable connectivity, federated learning remains vulnerable to gradient leakage attacks. Although existing defense schemes can provide some protection, they often lead to significant degradation in model performance or increased communication overhead.

    Addressing the above challenges, Professor Cui Haixia's research team proposed a risk-aware federated learning framework for edge scenarios. On one hand, the framework conducts privacy risk assessment based on the heterogeneous training configurations of edge devices and the gradient norms of model updates, subsequently activating adaptive defenses. Specifically, by combining subtractive dithering quantization, it injects controllable Gaussian noise while quantizing and compressing model data, achieving reliable privacy protection and reducing communication costs by over 50%. On the other hand, the framework employs a noise-aware aggregation strategy, dynamically adjusting client weights to mitigate the negative impact of noisy updates, thereby achieving joint optimization among privacy, communication efficiency, and model accuracy. In summary, the proposed training framework balances data security and communication overhead optimization, demonstrating significant potential for general model training tasks on heterogeneous edge devices.

     This work was supported by the National Natural Science Foundation of China, the Natural Science Foundation of Guangdong Province, and other projects.

    Link: https://url.scnu.edu.cn/record/view/index.html?key=7422ddd7239bc7d1345d0b1f3ad33091