Homomorphic threshold detection for wireless sensor data using a privacy-preserving machine learning model

There are several privacy and security issues when public clouds are used to host sensitive data from a Wireless Sensor Network (WSN). Traditional encryption procedures provide the required data security while it is being transmitted and stored. However, mathematical data processing on encrypted data cannot be carried out directly in public clouds. In such a scenario, Homomorphic Encryption allows privacy-preserving mathematical operations among the stored encrypted operands without decryption at the Public Cloud Server (CS) where the available powerful processing capabilities of the CS are efficiently utilized. In this paper, Threshold Detection (TD) is carried out on the homomorphically encrypted WSN data using a Privacy-Preserving Machine Learning (PPML) model. The PPML model is trained, and tested at the Base Station (BS) of the WSN. Then, the trained/tested PPML model is deployed at CS for online aggregate services like SUM, AVG, MIN, MAX, COUNT, SORT, SEARCH and TD. The model selected for PPML is a Feed-Forward Neural Network(FFNN) for Classification which is well-suited since the response is the binary outcome (0 or 1). Randomized encryption is used to prevent the Chosen Plaintext Attack (CPA). Several challenges posed by PPML are solved ingeniously by taking care of accuracy and scalability. Simulation results in terms of appropriate metrics are presented. It is experimentally found that a classification accuracy of 99.999% is reached after fine-tuning the hyperparameters of the custom-built FFNN classifier.