A Federated Incremental Approach For Distributed Data Security With Random Weight Neural Networks

As digital technology enters a period of rapid development,technologies such as big data and artificial intelligence are experiencing an explosion,and with this comes the issue of data privacy and security.Today,there is an increasing focus on data privacy and security,and as a result,countries are strengthening their protection of data security and privacy with the establishment of various regulations.The establishment of these regulations has,to varying degrees,posed new challenges to AI’s traditional data processing model.Federated learning was pioneered by Google Research in 2016 as one of the solutions to the current dilemma of big data,where traditional methods alone have become a bottleneck.The mechanism that the training data remains locally stored in the participants during the federated learning process enables the sharing of training data among participants while ensuring the protection of each participant’s privacy.As a result,federated learning has gained popularity among scholars.However,most of the traditional federated learning algorithms are implemented based on gradient algorithms,which have insurmountable technical bottlenecks such as slow convergence speed,easy to fall into local minima,and strong dependence on the setting of initial parameters.As a lightweight and fast shallow neural network,the introduction of random weight incremental neural network into federated learning will be an effective solution to the appeal problem.However,its randomness and incremental learning characteristics will bring many challenging problems to the implementation of federated learning,so this thesis carries out research on the federated incremental learning method of random weight neural network for big data security,and the main innovative work is summarised as follows:(1)A Horizontal Federated Incremental Learning Method with Random Weight Neural Network(HFIRWNN)is proposed to realize distributed collaborative modeling under data privacy security by constructing a random weight neural network as the edge-end model for the application scenario where the datasets at the edge-end have the same feature space and different sample spaces.(2)A Vertical Federated Incremental Learning Method with Random Weight Neural Network(VFIRWNN)is proposed to address the application scenario where the data sets at the edge ends have the same sample space and different feature The new vertical federated framework is built by introducing the privileged information paradigm into the federated learning,which solves the problem of interdependence of the edge ends in the traditional vertical federated framework and makes it difficult to be applied independently.(3)To improve the learning efficiency of the edge-end random model,stochastic configuration networks are introduced and Federated Incremental Learning Method with Stochastic Configured Networks(FSCNs)based on stochastic configuration networks is proposed.For transversal federated learning,a Weighted Aggregate Strategy(WA)and a Greedy Choice Strategy(GC)are constructed for the hidden layer nodes,and each strategy is designed with a speed priority algorithm and a quality priority algorithm,respectively.algorithm and quality priority algorithm respectively.Vertical Federated Learning SCNs(VFSCNs)are proposed for vertical federated learning by combining a supervisory mechanism with a privileged information paradigm,which addresses the shortcoming that the traditional vertical federated framework must rely on collaborators for joint prediction.In addition,HFSCNs and VFSCNs address the problem that some of the HFIRWNN and VFIRWNN hidden layer nodes have no or little contribution to the model.In summary,to cope with the collaborative modelling problem under data privacy security,this thesis focuses on horizontal federated incremental learning and vertical federated incremental learning methods using random weight neural networks as edgeend learning,and constructs different learning algorithms for different real-life scenarios to develop federated learning into an incremental learning model.