The advent of IoT devices led to the creation of many different, distributed machine learning techniques to identify patterns and make logical decisions based on the large volumes of data generated. Traditionally, these devices readily shared their raw data with a centralised server. However, this sharing of raw data is not conducive to maintaining the privacy of the data. One way that this has been overcome is through the use of Federated Learning. This reduces the volume of data shared, as only the parameters obtained from locally trained models are forwarded to the central server. However, this does not fully overcome the security and privacy concerns.
Somewhat homomorphic encryption (SHE) schemes allow computations to be performed on encrypted data without needing to decrypt it first. One such scheme is CKKS. It approximates computations on encrypted floating-point numbers, and so it is an attractive candidate for use for carrying out privacy-preserving numerical computations.
This project will consider the use of somewhat homomorphic encryption alongside federated learning – exploring its feasibility for deployment in real-world IoT settings.
You need to be comfortable working with maths, networking and programming to undertake this project.