Final Project Report of IOTC Pilot Grant Close Out Protect Privacy in a Distributed Learning Platform with a Natural Language Processing

Grant Project Final Report for Completion/Outcomes for: The goal of this project is to accelerate research for privacy-sensitive IoT applications by using mobile platform as a service (PaaS) model and data hub. Machine Learning as a Service (MLaaS) was designed in a common network communication architecture for collaborative deep learning-based training and automatically model deployment and Database as a Service (DaaS) is designed to provide user access and data privacy control. A supervised distributed deep learning system is prototyped in this pilot study. The system is built on end devices to explore the feasibility of delivering distributed deep learning models in a mobile environment. This collaborative learning scheme enables multiple mobile devices to train a global model together without releasing privacy-sensitive user data but only sharing partial neural network parameters among end devices. The PIs has developed an approach that encrypts the entire database of user ids and passwords with theoretical solutions such as semantic preserving encryption techniques and/or fully homomorphic encryption, which allow the cloud servers 1) to compute queries over the encrypted data without decrypting at any point of the communication. Only the encrypted database is kept on the untrusted cloud server at any time of the communication and only clients see decrypted data after the communication is done and 2) to compute updating parameters of deep learning models of the cloud ML server directly on the encrypted parameters with user sent parameters in encrypted which were generated from training with user own data on users’ mobile phones. This design achieves privacy features with a proxy server on the distributed deep learning systems with semantic-preserving encryptions and retrieval techniques that allows cloud servers blindly retrieving a user id and password directly from the encrypted database without decrypting process over the entire communication routes while the encryption is transparent to the users, and the database server on cloud is not aware of the encryption. This design has been applied to smoking cessation application using NLP techniques and wearable IoT sensor data processing.