Lightweight Homomorphic Encryption for Privacy-Preserving Machine Learning on Edge Devices

The increasing use of machine learning (ML) on edge devices, including smartphones, IoT sensors, and wearable devices equates significant privacy considerations, especially when users provided sensitive data to the cloud for further process. The proposed solution is a privacy-preserving machine learning (PPML) architecture that combines lightweight Homomorphic Encryption (HE) to provide a secure but efficient ML inference route on edge devices. The architecture will consist of three base modules: Edge Devices; an Encryption Module; and a Remote/-Cloud Server. The Edge Devices, when presented with user data, will primarily the locations in which data is collected and the local preprocessing occurs, where the abbreviated data reduces the data size and protect the associated privacy, before encryption occurs. Privacy-preserving systems often employ lightweight HE schemes, such as CKKS and Paillier, with beneficial traits of fast encryption that is current suitable for real-time edge applications with privacy in secured data. The Encryption Module used optimized lightweight HE schemes, including CKKS and Paillier, that enable rapid encryption suitable to real-time edge application environments, while keeping the data secure. Encrypted data is sent to a remote server for inference which protects the data privacy because the remote server never sees the raw data. The experimental evaluations of the proposed lightweight HE framework were on realistic edge hardware and achieved 91.8% accuracy, with large reductions in inference latency (220 ms), memory usage (85 MB), and energy usage (40 mWh) compared to traditional HE. Overall, it was established that the system achieves the necessary balance between data privacy and computational capability to be used in the real world environments for the privacy and resource-constrained edge systems.