CRII: III: Federated Learning with Enhanced Efficiency and Privacy Preservation over Heterogeneous Devices

Federated systems are currently popular because they provide computing resources over a distributed network. Examples of these systems include distributed cloud computing and mobile cellular networks. To enhance real-time decision-making, reduce latency, and address growing privacy concerns, federated learning (FL) has emerged as a paradigm for implementing artificial intelligence (AI) models at the edge. FL enables edge devices to collaboratively train a global model without sharing raw data. However, the FL framework encounters critical challenges: (1) the varied computational capabilities of edge devices cause inconsistent training convergence and potential latency issues; (2) the diverse data distributions across devices introduce biases, adversely affecting model accuracy; and (3) although raw data is not shared, model updates communicated among edge devices are still vulnerable to privacy breaches. This project aims to develop an innovative privacy-preserving FL framework suitable for heterogeneous edge devices without compromising performance. The anticipated outcomes of this project promise significant societal benefits by enabling new applications in various domains, such as healthcare, finance, and cybersecurity, where data privacy is paramount. Additionally, the investigator is committed to promoting educational diversity and empowering underrepresented groups in science and engineering fields. This commitment extends to fostering a multidisciplinary research environment for college students and broadening the interest of secondary school students in AI-related subjects. To address the three challenges mentioned, the proposed FL framework is divided into three phases: local training, data transmission, and server-based global model training. (1) During the local training phase, instead of deploying identical models to all edge devices, tailored device-specific models are locally trained. These local models are adaptive to the diverse computational capabilities and data distributions to mitigate biases and enhance model performance. (2) In the data transmission phase, differential privacy technology is utilized to protect data privacy on edge devices to prevent inference attacks. (3) The global training phase leverages the underutilized computational resources of the server, aiming to significantly improve the efficiency and performance of the global model. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.