通過消除用戶資料特徵衝突來改進特徵非獨立同分布的聯合學習

Abstract

聯邦學習可以在邊緣設備上協作訓練深度神經網絡，而無需集中數據以保護數據隱私。以往聯邦學習假設資料是獨立同分佈（iid），但該假設在現實世界中並不成立，因為客戶端的數據是異構的，我們稱這個現象為非獨立同分布 (non-iid)。在這種情況下，傳統聯邦學習框架的性能可能會因為數據分佈的不同而有顯著改變。最近的研究主要集中在標籤上的非獨立同分佈，其數據是基於標籤分佈進行分配。與此設定不同，我們將要處理一個更貼近現實的情形，即數據是基於特徵的異質性分配，造成這個行情的原因可能是不同的傳感器、不同的城市、不同的季節。我們稱這種情況為領域偏移非獨立同分佈(feature shift non-iid)。在這項研究中，我們提出 FedADA 通過消除領域衝突來解決這個問題。我們提出的方法在廣泛的實驗中優於傳統的聯邦學習框架以及現有最先進的資料增強方法。

Federated learning(FL) enables training deep neural networks collaboratively on edge devices without centralizing the data as well as preserving data privacy.The previous FL assumption of independent and identical distribution (iid) does not hold in the real world as the data on the client sides are heterogeneous, which has been called non-iid. Under this scenario, the performance of conventional federated learning frameworks may vary significantly according to the data distribution. Recent works focus on label non-iid, where data are distributed based on labels. Unlike this setting, we deal with a more realistic problem that data are heterogeneous based on the feature, e.g., different sensors, different cities, and different seasons. This scenario is viewed as a feature-shift non-iid that clients contain non-identical domain data. In this work, we propose FedADA that address this problem by mitigating the domain conflicts. Our proposed method outperforms both classical FL frameworks as well as the SOTA augmentation methods in extensive experiments. The empirical results are conducted with various datasets and model architectures to guarantee the performance of our approach.