| In recent years,federated learning(FL)has attracted widespread attention as a new AI edge computing paradigm.FL utilizes multiple edge devices as participants to collaboratively train the global model in deep learning,and the training process does not leak the local data of any participant.In a typical edge computing scenario,the participants in FL are usually heterogeneous and consist of various devices such as personal computers,smartphones,IoT devices,and network devices.In the heterogeneous scenario,devices with weak computing power are called stragglers,which will become the bottleneck of each training round in FL,thereby limiting the overall model convergence speed and global model accuracy.To address the laggard effects of the stragglers,a feasible solution is to outsource the stragglers’computing tasks to high-performance devices.The process of outsourcing is bound to transmit the original data of the stragglers through the network,which cannot meet the needs of privacy protection.Therefore,this dissertation proposes an efficient linear coding-based heterogeneous FL system framework to accelerate the model convergence speed in the context of heterogeneous federated learning and protect the data privacy of the participants.For the above-mentioned privacy problems and the problem of training speed reduction caused by stragglers,this dissertation studies the problems from the following two aspects:(1)In the scenario where a single straggler is included in a training round of FL,the task outsourcing for single straggler(TOSS)algorithm is proposed.The TOSS algorithm can select appropriate collaborative clients and outsource the tasks of the straggler to collaborative clients in one training round to speed up training.At the same time,a linear coding computing(LCC)scheme is designed to ensure the privacy protection in the TOSS algorithm.(2)In the scenario where multiple stragglers are included in a training round of FL,a multiple stragglers task outsourcing(MSTO)algorithm is proposed.The MSTO algorithm consists of two algorithms,the client-based multiple stragglers task outsourcing(C-MSTO)and the server-based multiple stragglers task outsourcing(S-MSTO).Meanwhile,a linear progressive coding(LPC)scheme is proposed to protect the privacy of outsourced data and further reduce the time consumption of encoding and decoding operations.This dissertation builds a generic FL system platform based on Python,and conducts a large number of experiments to evaluate the proposed linear coding FL framework from different perspectives.The experimental results show that when the participating devices of FL have large differences in computing power,the proposed linear coding FL scheme can reduce the system training time by more than 90%compared with traditional FL. |
