Game of Gradients: Mitigating Irrelevant clients in Federated Learning
AAAI Conference on Artificial Intelligence
Though FL paradigm has received significant interest recently from the research community, the problem of selecting the relevant clients w.r.t. the central server’s learning objective is under-explored. We refer to these problems as Federated Relevant Client Selection (FRCS). Because the server doesn’t have explicit control over the nature of data possessed by each client, the problem of selecting relevant clients is significantly complex in FL settings.
The problems of FRCS need to be resolved:
- selecting clients with relevant data
- detecting clients that possess data relevant to a particular target label
- rectifying corrupted data samples of individual clients.
We follow a principled approach to address the above FRCS problems and develop a new federated learning method using the Shapley value concept from cooperative game theory. Towards this end, we propose a cooperative game involving the gradients shared by the clients. Using this game, we compute Shapley values of clients and then present Shapley value based Federated Averaging (S-FedAvg) algorithm that empowers the server to select relevant clients with high probability.
Key Words:Relevant data,Shapley value,Federated Learning
FL setting assumes a syn- chronous update of the central model and the following steps proceed in each round of the learning process (McMahan et al. 2017).
Note that only a fraction of clients is selected in each round as adding more clients would lead to diminishing returns beyond a certain point.
Though initially the emphasis was on mobile- centric FL applications involving thousands of clients, recently there is significant interest in enterprise driven FL applications that involve only a few tens of clients.
We apply standard FedAvg algorithm to two cases:
(a) where all clients possess relevant data
(b) where some clients possess irrelevant data.
To simulate irrelevance, we work with open-set label noise (Wang et al. 2018) strategy, wherein we randomly flip each odd label of the 4 irrelevant clients to one of the even labels.
Contributions of Our Work:
0x02 Related Work
0x03 Problem Statement
0x04 Proposed Solution
We posit that more the relevance value for a client, more is its likelihood to be relevant and thereby more is its contribution to the objective of central model.
- Extreme 1-class-non-iid
As is typical with the federated learning setting, we assume that the data is distributed non-iid with each client. We follow the extreme 1-class-non-iid approach mentioned in (Zhao et al. 2018b) while distributing the data to clients.
2.S-FedAvg: To Selects Relevant Clients
3.Class-specific Best Client Selection
4.Data Label Standardization