New Jersey Institute of Technology, USA
Topic: Federated Learning for Mobile Sensing Data
Abstract: Federated Learning (FL) has emerged as a new distributed machine learning paradigm that enables privacy-aware training and inference on mobile devices with help from the cloud. FL has the potential to enable a wide range of new mobile apps that benefit from running machine learning models on mobile sensing data. The privacy-sensitive raw data is used for local training on the devices, and only the model parameters are transferred to the cloud, where a global model is aggregated and shared with all mobile devices. This keynote talk presents our ongoing work on FL systems and applications. First, we describe FLSys, an end-to-end FL system designed to achieve energy efficiency, tolerance to communication failures, and scalability. In addition, different FL models, accessed concurrently by different apps, are able to work with different FL aggregation methods in the cloud. A common API is provided for third-party app developers to train FL models. FLSys is implemented in Android and AWS cloud. We demonstrate FLSys in the context of human activity recognition (HAR) in the wild, with data collected from the phones of 100+ students. We propose a novel HAR-Wild model, which is based on a skipped Convolution Neural Network model with a data augmentation mechanism to mitigate the non-Independent and Identically Distributed data problem that negatively affects FL model training. We conduct extensive experiments on Android phones and Android emulators, showing that FLSys and HAR-Wild achieve good model utility and practical system performance, in terms of training time and resource consumption on the phones. Second, we present a system for fine-grained location prediction (FGLP) of mobile users, based on GPS traces collected on the phones. FGLP has two components: an FL framework and a prediction model. The framework runs on the phones of the users and also on a server that coordinates learning from all users in the system. The framework represents the user location data as relative points in an abstract 2D space, which enables learning across different physical spaces. The model merges Bidirectional Long Short-Term Memory (BiLSTM) and Convolutional Neural Networks (CNN), where BiLSTM learns the speed and direction of the mobile users, and CNN learns information such as user movement preferences. Our experimental results, using a dataset with over 600,000 users, demonstrate that FGLP outperforms baseline models in terms of prediction accuracy for pedestrians and bicyclists. In addition, benchmark results on several types of Android phones demonstrate FGLP’s feasibility in real-life. We conclude this talk with lessons learned from building FL systems and applications, and with challenges that still need to be overcome in order to deploy FL models in real-life.
Bio: Cristian Borcea is a Professor of Computer Science and the Associate Dean for Strategic Initiatives in the Ying Wu College of Computing at New Jersey Institute of Technology, USA. He also holds a Visiting Professor appointment at National Institute of Informatics, Tokyo, Japan. Cristian has over 20 years of experience in the fields of mobile computing & sensing; ad hoc & vehicular networks; and cloud & distributed systems. His current research is at the intersection of mobile computing and machine learning. He has published over 100 papers in top international journals and conferences, and his research has been covered in over 20 media articles in the past few years. Cristian has served as Technical Program Chair or General Chair to conferences such as IEEE MDM, IEEE Mobile Cloud, and EAI Mobiquitous. Cristian received his PhD in Computer Science from Rutgers University, USA. More information: http://cs.njit.edu/~borcea.