Toward a Federated Model for Human Context Recognition on Edge Devices

  • Wandemberg Gibaut Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Alexandre Osorio Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Amparo Munõz Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Sildolfo F. G. Neto Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Daniel Miranda Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Felipe Santos Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Michelle Scarassati Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
  • Fabio Grassiotto Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC), Eldorado Research Institute, Campinas- SP
Keywords: Artificial Intelligence, Fuzzy and neural systems relevant to control and identification, Human Activity Recognition, Federated Learning

Abstract

Federated Learning is a promising technology to address crucial problems, such as those related to data privacy, involved in training Machine Learning (ML)/Deep Learning (DL) models in a distributed way. On the other hand, Human Activity Recognition (HAR) has recently gained more attention due to the evolution of the technologies involved, such as sensor availability, advances in ML/DL/Edge AI, and IoT. Due to computational resource constraints, techniques must be employed to reduce the effort required to train the model on the device. Meanwhile, there’s the need to customize the ML model of each Federated Learning (FL) client with the specific data collected by that client. The present work explores the FL of an ML model for HAR in a set of twelve simulated FL clients, each with its own set of data from smartphone sensors. The FL loop starts from a global model that was previously trained in a centralized way, using a large dataset, different from the data used individually by each client during the FL. In this way, the FL constitutes a fine-tuning of the base model. The metrics collected are balanced accuracy and loss. Data is pulled from the ExtraSensory dataset, creating a benchmark for future applications across device farms and in-the-wild devices. The results show that our models achieve equivalent or better performance than most methods found in the literature, using a relatively simple Multilayer Perceptron (MLP) model. The proposed method can then reduce the time needed to retrain the model when data is acquired from the device’s own sensors.
Published
2022-10-19
Section
Articles