Device-to-Device Collaborative Learning for Self-Localization with Previous Model Utilization

As AI technology is becoming to be used in various fields, federated learning (FL) has been attracting attention in terms of security for training data. Especially, fully-decentralized FL is expected to be applied for tasks that multiple devices or vendors participate in because of its edge device learning capabilities. This paper focuses on wireless ad hoc federated learning (WAFL), which is one of these fully-decentralized FL without client-server communication over the Internet, and proposes an adaptation of WAFL for photographer location classification system (WAFL-PLCS). In addition, we study the mixture of the previous model parameters during training to improve accuracy. WAFL-PLCS achieved about 85% accuracy when there is no bias in the distribution of training data labels which each device utilize. By contrast, in the unbalanced case, WAFL-PLCS showed about 80% accuracy, both with and without previous model aggregation. This indicates that photographs can be utilized to estimate an approximate location at a certain level of accuracy, and such a system can be implemented without the need to share the photographs themselves.