Federated Transfer Learning with Multimodal Data

Abstract

Smart cars, smartphones and other devices in the Internet of Things (IoT), which usually have more than one sensors, produce multimodal data. Federated Learning supports collecting a wealth of multimodal data from different devices without sharing raw data. Transfer Learning methods help transfer knowledge from some devices to others. Federated Transfer Learning methods benefit both Federated Learning and Transfer Learning. This newly proposed Federated Transfer Learning framework aims at connecting data islands with privacy protection. Our construction is based on Federated Learning and Transfer Learning. Compared with previous Federated Transfer Learnings, where each user should have data with identical modalities (either all unimodal or all multimodal), our new framework is more generic, it allows a hybrid distribution of user data. The core strategy is to use two different but inherently connected training methods for our two types of users. Supervised Learning is adopted for users with only unimodal data (Type 1), while Self-Supervised Learning is applied to user with multimodal data (Type 2) for both the feature of each modality and the connection between them. This connection knowledge of Type 2 will help Type 1 in later stages of training. Training in the new framework can be divided in three steps. In the first step, users who have data with the identical modalities are grouped together. For example, user with only sound signals are in group one, and those with only images are in group two, and users with multimodal data are in group three, and so on. In the second step, Federated Learning is executed within the groups, where Supervised Learning and Self-Supervised Learning are used depending on the group's nature. Most of the Transfer Learning happens in the third step, where the related parts in the network obtained from the previous steps are aggregated (federated).