FedSiM: a similarity metric federal learning mechanism based on stimulus response method with Non-IID data

Abstract

Abstract Federal learning based on parameter sharing under the assumption that the data obey independent identical distribution (IID has already achieved good results in areas such as fault diagnosis. Data collected by the decentralized devices often do not obey IID. However, when faced with the scenario of client data obeying Non-IID distribution, its diagnostic accuracy is usually weak. Based on this, we did an investigation on the mechanism causing this phenomenon and found that it was attributed to the weight shift of the network. Therefore, based on the stimulus response principle, we investigated the network similarity of federal clients under different data distributions and explain the reasons for the weight shift. Firstly, it was pointed out that there are differences in the regions where the network is activated when performing different classification tasks. Then, similarity metric federal learning (FedSiM) was proposed based on the principle that there are differences between the activated regions. Finally, experiments were designed on the Case Western Reserve University bearing failure dataset for different degrees of IID cases. The results show that FedSiM can improve the diagnostic accuracy by 15.8 percentage points in the case of Non-IID, and a few shared FedSiM methods to further improve the accuracy were also given.