Abstract
Copyright © 2025 Inderscience Enterprises Ltd.Many fields have seen success with deep learning implementations, including bioinformatics, photo processing, gaming, computer security, etc. However, a large amount of training data is typically required for deep learning, which may not be made available by a single owner. As the amount of data continues to rise at an exponential rate, many people are turning to remote cloud services to store their information. Human activity recognition (HAR) provides massive amounts of data from IoT devices to collaboratively construct predictive models for medical diagnosis. To protect users’ anonymity in scenarios where DNNs are used in HAR learning, we present Multi-Scheme Differential Privacy. MSDP uses a multi-party, secure variant of the ReLU function to cut down on transmission and processing time. MSDP is proved to be secure in comparison to existing state-of-the-art models without compromising privacy through experimental validation on four of the most popular human activity detection datasets.
NETL
NSTL
Inderscience Enterprises Ltd