Recognizing human activities in a privacy-preserving way

Human Activity Recognition (HAR) is a classification problem that aims to discern well-defined human activities through sensor data. This thesis explores HAR with a focus on privacy concerns by adopting the PHAR architecture. The architecture adopts a decentralized approach, moving the primary computational effort to the data. Additionally, the purpose of the project is to illustrate how nowadays smartphone devices can be used to collect and train on models their own sensor data. The architecture for this project is built on top of a Flower framework with Android devices connected to a server that performs Federated Learning (FL) of convolutional neural network (CNN) exploiting federated averaging (FedAvg). The devices are loaded with Tensorflow lite (TLite) quantized pretrained CNN models using the technique of Transfer Learning (TL). The results show that mitigating the privacy in edge devices don't affect considerably the model metrics results instead the scalable architecture can potentially be used to generalize the model of the resilient network, getting rid of the collection of a large amount of data and reaching a local convergence fastly. Despite these findings, privacy could potentially be compromised delegating the issue to the weights integrity or authentication of clients. Finally, this study establishes a foundation for future research with the potential to integrate real time HAR on smartphones as well as making it portable to other OS devices.