Qualcomm Collaboration with Ying Wu College of Computing Brings Privacy-Preserving Machine Learning

Imagine a digital system that could predict physical and mental health conditions among college students based on early symptoms and quickly connect them to counseling and psychological services on campus. While a seemingly obvious application of IoT (Internet of Things) technology, the implementation of such a system would need to avoid exposing and storing sensitive health data along the way, which could run afoul of relevant federal laws and compromise user privacy.  A system developed by NJIT researchers and Qualcomm Technologies Inc. is now able to do this by ensuring that the privacy-sensitive student data never leaves the end-users' devices during the entire process and all raw data processing is done on the devices. This is achieved using federated learning, which runs a decentralized AI model between the end-user devices and the cloud, without the need to share user data with the cloud.

A three-year collaboration between Qualcomm Technologies, a multinational wireless technology corporation, and Ying Wu College of Computing’s Cristian Borcea and Hai Phan (faculty in the departments of computer science and data science respectively) has led to the inventions of Federated Learning System (FLSys) and Zone Federated Learning (ZoneFL), the first end-to-end, mobile-cloud federated learning (FL) systems that work effectively on smart phones. Federated Learning System provides practical, efficient, scalable and privacy-preserving FL for data collected on mobile devices. Zone Federated Learning is the team’s next-level enhancement that adapts FL models to mobility behavior in different geographical zones. Both findings have the potential to transform how user-based mobile information, such as accelerometer, heart rate, etc., are leveraged in novel applications of machine learning, while protecting the privacy of user data. Descriptions of these systems were recently published in IEEE Transactions on Mobile Computing and presented at the 21^st IEEE International Conference on Pervasive Computing and Communications (PerCom 2023).

The collaboration between NJIT and Qualcomm Technologies started with a serendipitous meeting between Prof. Phan and Dr. An Chen, VP of Engineering at Qualcomm Incorporated at a conference in Chicago in 2019. Phan’s research in human behavior modeling and trustworthy AI using wearable devices sparked Chen’s interest and subsequent conversations. Talking about this research collaboration, Dr. Chen says: “Qualcomm Technologies aims to harvest the power of billions of mobile and IoT devices to develop novel AI-driven products at the network edge. We chose to collaborate with Phan and Borcea due to their expertise in machine learning and mobile computing. Furthermore, NJIT’s faculty have demonstrated the ability to seamlessly integrate fundamental research into practical system development, which could lead to commercial products.”

Federated learning uses machine learning techniques to train an algorithm across multiple decentralized edge devices or servers holding local data samples, without exchanging them, to make predictions or classifications.  Practical difficulties arise, however, when trying to make federated learning efficient on resource constrained mobile devices.

The initial collaboration between NJIT and Qualcomm Technologies, which included several NJIT PhD students, resulted in Federated Learning System, a system designed specifically for resource constrained mobile devices on which many other machine learning applications can be built. Federated Learning System enables the creation of an ecosystem for app developers that makes it possible for [almost] anybody to innovate in the area of machine learning for better health and wellness, better traffic prediction to avoid congestions, or whatever else will improve the quality of life for society - “with a little bit of skill!” Borcea added, lest one thinks his invention is nothing more than an enabler for machine learning DIY.

Zone Federated Learning takes the Federated Learning System concept one step further by dividing the physical space into location-based zones. This smart improvement will not only respond to a larger mass audience, but one from which data compiled in diverse urban and rural areas where climate, population size and economic standards can determine the quality of the machine learning predictions. Borcea said, “Behavior depends on geography. Now, we can predict things such as heart rate while running or hiking among a broad spectrum of individuals in different geographical regions.”

To evaluate the research, field studies were conducted with over 100 NJIT students collecting data on their phones “in the wild” (outside a lab) as they went about their daily routines. The results demonstrated the feasibility of the solutions in real-life situations, with further improvements increasing speed, extending battery life, and perfecting a lightweight design.

The NJIT-Qualcomm Technologies collaboration results not only in top notch science, but also in software prototypes that allow for technology transfer between academia and industry.

Future research will focus on making federated learning practical on IoT devices, which have even more stringent resource constraints than mobile phones.