Client selection based weighted federated few-shot learning

With the advancement of technology, clients have a large amount of personal data. Human beings are paying more and more attention to the privacy and security of this part of data. Clients do not want to share private data, which directly leads to the existence of data islands. Especially in few-shot scenarios, due to the insufficient amount of personal data, constructing an effective few-shot model is difficult. To solve the above problems, we propose Federated Few-shot Learning (FedFSL) in this paper. We utilize Federated Learning (FedL) to ensure the privacy and security issues of joint training. Moreover, the global model obtained by FedL has the characteristics of universally applicable to all clients. That universality satisfies the scenario of universally applicable to all meta tasks in the fewshot meta-learning stage. What is more, to obtain a more effective global federated universal few-shot model, we respectively proposed Weighted FedL (WFedL) and Client Selection based FedL (CSFedL) strategies. WFedL takes into account the difference between clients performance when building the global model and assigns different weights to different clients. CSFedL considers the malicious participation of clients, and we propose an adaptive client selection strategy to mitigate the impact caused by malicious participation. Extensive federated experiments on CIFAR-10 and CIFAR-100 show the advantage of proposed WFedL, CSFedL and combined Client Selection based WFedL (CSWFedL). We further verify the performance improvement of FedFSL on miniImagenet and propose our overall framework Client Selection based WFedFSL (CSWFedFSL). The best performance of CSWFedFSL is higher than both the few-shot baseline and FedFSL, and CSWFedFSL protects clients data privacy in the few-shot scenario. (c) 2022 Elsevier B.V. All rights reserved.