[4Yin2] Interactive session 2

[4Yin2-26] Federated Learning using non-overlapping features by semantic encoding

〇Takeshi Mori¹, Hirofumi Kondo¹, Takuji Nagao² (1.The Japan Research Institute, Limited, 2.Sumitomo Mitsui Banking Corporation)

Keywords:Federated Learning, Machine Learning, Privacy Preserving, PETs

Federated Learning (FL) has attracted much attention as Privacy-Enhancing Technologies (PETs) that enable machine learning with distributed data among multiple organizations without exchanging them. Horizontal Federated Learning (HFL) is a method of FL, which is applicable when organizations participating in FL computation have the same features for different samples. However, in the real world, features held by multiple organizations rarely overlap perfectly. Furthermore, if HFL is applied when there are few features that overlap among organizations, the performance improvement will be limited. In this paper, we propose methods to improve the performance of models trained by HFL when the number of overlapping features among organizations is limited. Our proposed methods transform non-overlapping feature values into other values that have the same meaning among organizations and use them as overlapping features of HFL. Experiment results show that our proposed methods improved the performance of HFL by combining the overlapping features with transformed features.