Perspectives on the Feasibility and Adoption of Fully Homomorphic Encryption - A Fully Homomorphic Encryption Application on Spam Detection with Machine Learning Models

Abstract

The general adoption of cloud computing resulted in frequent security incidents, including data breaches in both the public and private sectors. Cloud service providers operate without transparency, fail to comply with regulations, and exploit private data for targeted advertising and training artificial intelligence models. The major tech companies have also been found to collaborate with intelligence agencies to illegally surveil individuals and governments by sharing private data, such as email communications. This study analyses the feasibility of Fully Homomorphic Encryption (FHE) as a solution to these security and privacy concerns, focusing on spam detection and email providers as representative subjects in the universe of outsourced computations on users’ private data. FHE maintains data privacy by enabling computations while the data is still encrypted, but it demands substantial computing and memory resources. The development of FHE-based applications is complex, requiring advanced knowledge of mathematical and cryptographic concepts. This work assesses the feasibility of FHE through experiments in spam detection by implementing Fully Homomorphic Encryption Spam Detector (FHE-SD), an application using the Concrete-ML libraries, which abstracts the complexity of FHE and simplifies its adoption. The experimental environment is a device with limited hardware resources, chosen to test if FHE can function without specialized hardware. For meaningful results, FHE-SD supports spam detection using machine learning algorithms, which are commonly used for spam detection. Four machine learning models are implemented in FHE-SD, in their FHE and on-clear versions, enabling various metrics and performance comparisons against traditional approaches.