YLYutao Liu

Yutao Liu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Yutao Liu has a total of 11 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Adaptive Model Pruning and Knowledge Codistillation for Heterogeneous Federated Learning

    by Zhihao Zeng, Xiaoning Zhang, Lang Fan, Yutao Liu, Ahmed Zoha, Muhammad Ali Imran

    Federated Learning (FL) enables edge devices to collaboratively train a shared model without exposing their local data. However, system heterogeneity in FL often prevents resource-constrained edge devices from training large models, resulting in the global model missing important data features. Simultaneously, statistical heterogeneity can lead to model divergence, further impairing global model performance. In this paper, we propose FedPKD, a unified framework that integrates Pruning and Knowledge coDistillation to address both system and statistical heterogeneity in FL. Specifically, FedPKD employs a two-stage pruning strategy. The server first constructs a set of candidate sub-models with different levels. Then the server assigns corresponding sub-models to devices based on their dynamic resource conditions. This design enables devices to train effectively under heterogeneous resource constraints, while ensuring stable aggregation of the sub-models. FedPKD also adopts the codistillation method, where local sub-models enhance their generalization ability by distilling knowledge from the most recent global sub-model. This helps address the performance degradation caused by Non-IID data. Finally, extensive experiments show that FedPKD can efficiently train sub-models under clients’ resource constraints, saving both computation and communication overhead while achieving competitive accuracy in Non-IID settings.

  • Accelerating Communication for UAV-Enabled Federated Learning With Adaptive Routing and Robust Aggregation Over Edge Networks

    by Yutao Liu, Xiaoning Zhang, Zhihao Zeng

    Nowadays, manufacturers are beginning to apply Artificial Intelligence (AI) technologies to Unmanned Aerial Vehicles (UAVs) for enhanced data processing, target recognition, tracking capabilities, etc. Federated Learning (FL) is a promising distributed machine learning technique that can perform these above complicated functions by providing UAVs with powerful learning capabilities. Since the dynamics of UAVs, parameter aggregation for UAV-enabled FL is accomplished in ground edge networks. However, traffic flows of the ground edge network are always changed and unpredictable due to the dynamics of accessed UAVs. It slows down the process of training traffic communication of UAV FL and thus hampers the model knowledge sharing. Things would get worse with a fixed FL central parameter server in the ground edge network when the location of UAVs may change at any time. To address these issues, we investigate the joint optimization problem of FL training traffic steering and dynamic central parameter server selection in ground edge networks in this study. Specifically, we leverage a Deep Reinforcement Learning (DRL) routing method to adaptively steer training traffic for ground edge networks. Then, based on the DRL routing agent, the robust central parameter server selection is performed in the UAV-enabled FL training process. Extensive simulation results show that our proposed method significantly decreases the communication time in UAV-enabled FL training up to 65.9% compared with the benchmark algorithms.


  • FedEco: Achieving Energy-Efficient Federated Learning by Hyperparameter Adaptive Tuning

    by Yutao Liu, Xiaoning Zhang, Zhihao Zeng, Shui Yu

    In Federated Learning (FL), each participating client needs to frequently compute the local gradient updates and communicate with the central parameter server, which causes high energy consumption on clients. Such energy consumption brings a significant challenge to battery-constrained clients (such as mobile devices, notebooks, IoT sensors, etc.) for edge networks. To tackle this challenge, in this study, we propose a method named FedEco to minimize energy consumption of all participating clients in FL. FedEco has the following features: i). FedEco achieves energy consumption minimization through hyperparameter1 adaptive tuning, which means allocating several kinds of optimized hyperparameters to the FL clients periodically. ii). The used hyperparameters include computing hyperparameters (i.e., training speed and training volume) and communication hyperparameters (i.e., transmission speed and transmission volume). iii). FedEco can theoretically guarantee FL convergence while satisfying the heterogeneous and dynamic edge environments. iv). FedEco finds the optimal computing and communication hyperparameters by decomposing the studied problem into several subproblems and solving them iteratively. Therefore, FedEco can achieve a good tradeoff between energy efficiency and model training convergence. We conduct extensive simulations to show that FedEco significantly saves energy consumption by up to 87.04% compared with the benchmark algorithms on average while guaranteeing the convergence of model training.


  • Self-Adaptive Gradient Quantization for Geo-Distributed Machine Learning Over Heterogeneous and Dynamic Networks

    by Chenyu Fan, Xiaoning Zhang, Yangming Zhao, Yutao Liu, Shui Yu

    Geo-Distributed Machine Learning (Geo-DML) has been proposed to collaborate geographically dispersed data centers (DCs) and train large scale machine learning (ML) models for various applications. While Geo-DML can achieve excellent performance, it also injects massive data traffic into the Wide Area Networks (WANs) in order to exchange gradients during model training process. Such a huge amount of traffic will not only incur network congestion and prolong the training procedure, but also result in straggler problem when DCs are working in heterogeneous network environments. To alleviate these problems, we propose Self-Adaptive Gradient Quantization (SAGQ) for Geo-DML in this work. In SAGQ, each worker DC adopts specific quantization method based on the heterogeneous and dynamic link bandwidth in order to reduce the commu- nication overhead and balance the communication time among worker DCs. By doing so, SAGQ will speed up the Geo-DML training process without sacrificing the ML model performance. Extensive experiments show that compared with the state-of-the- art techniques, SAGQ reduces the Wall-clock time spent to train an ML model by 1.13× -21.31×. In addition, SAGQ can also improve the model accuracy by0.11%-2.27%over baselines.


  • Chronos: Accelerating Federated Learning With Resource Aware Training Volume Tuning at Network Edges

    by Yutao Liu, Xiaoning Zhang, Yangming Zhao, Yexiao He, Shui Yu, Kainan Zhu

    Due to the limited resources and data privacy issue, last decade witnesses the fast development of Distributed Machine Learning (DML) at network edges. Among all the existing DML paradigms, Federated Learning (FL) would be a promising one, since in FL, each client trains its local model without sharing the raw data with others. A community of clients with the same interest can join together to derive a high-performance model by periodically synchronizing the parameters of their local models under the help of a coordination server. However, FL will encounter the straggler problem at network edges, and hence the synchronization among clients becomes inefficient. It slows down the convergence speed of learning process. To alleviate the straggler problem, we propose a method named Chronos that accelerates FL with training volume tuning in this paper. More specifically, Chronos is a resource aware method that adaptively adjusts the amount of data used by each client for training ( i . e . training volume) in each iteration in order to eliminate the synchronization waiting time caused by the heterogeneous and dynamical computing and communication resources. In addition, we theoretically analyze the convergence of Chronos in a non-convex setting and utilize the results for the algorithm design of Chronos in return to guarantee the convergence. Extensive experiments show that compared with the benchmark algorithms ( i . e BSP and SSP), Chronos significantly improves convergence speed by up to 6.4×.