Decentralized Collaborative Learning Algorithms For Complex Network Systems

Decentralized collaborative learning,in which multiple participants independently train the learning model and communicate in the network in a P2P manner to train the network sharing model,is becoming a new trend to solve the problem of data scarcity.At the same time,with the development of artificial intelligence and communication technology,decentralized collaborative learning has been widely applied in many fields,such as the Internet of Things(IoT),edge computing,social networks,etc.However,the design of a decentralized collaborative learning algorithm faces great challenges due to the complexity of network systems.For example,the nodes in the network are constantly changing,and it is impossible to expect the moment when the network stops.With the continuous increase of network scale,the communication between nodes will be more frequent.The nodes in the network may have Byzantine faults or malicious nodes.In the network,the bandwidth is restricted and the communication between nodes is very limited.At the same time,although there is no raw data to share and no central node to coordinate global training in the communication process,frequent communication between nodes still inevitably leads to privacy leakage.Therefore,the design of efficient and highly robust decentralized collaborative learning algorithms for complex network systems has become an urgent problem to be solved.Based on the above complex network system model,the main contribution of this dissertation can be summarized as follows:(1)For the multi-hop dynamic network with Byzantine nodes,we design an efficient distributed uniform sampling algorithm.Based on MetropolisHastings Random Walk(MHRW),the algorithm achieves a uniform sampling by balancing the deviation caused by the inconsistent connectivity between different nodes.This algorithm requires lightweight time and communication complexity and it is an important tool for the design of large-scale dynamic network algorithms.(2)For reinforcement learning tasks,every individual in the network needs to choose one option with the greatest reward from the given options with unknown stochastic qualities,under the network with incomplete or even falsified information.Then they communicate with each other and finally choose the best option.By designing the three-stage learning(sampling,selection,adoption)algorithm,the communication between individuals can be effectively reduced.At the same time,the communication and time complexity can be kept lightweight under the premise of ensuring the same convergence with the algorithm based on full information.(3)For machine learning tasks,we design a dynamic decentralized parallel stochastic gradient descent algorithm(D-(DP)2SGD)based on differential privacy,which protects the data transmission during the learning process.Through rigorous analysis,the algorithm satisfies ?-differential privacy.Simultaneously,this algorithm reduces the convergence rate from O(1/(?))to O(1/(?))compared with the centralized learning.In this dissertation,we design decentralized collaborative learning algorithms,which are suitable for large-scale dynamic networks with fault nodes and the limited network communication,and the algorithms achieve the optimal convergence rate.Moreover,it can be widely applied to the field of Internet of Things,edge computing,social networks and so on.