Enhancing Privacy in Artificial Intelligence Services Using Hybrid Homomorphic Encryption

Abstract

The increasing occurrence of cyberattacks specifically aimed at critical infrastructure has led to the adoption of network intrusion detection techniques for the Internet of Things (IoT).  AI is transforming multiple sectors today, the growth of adversarial attacks on AI models and models present imperative privacy issues which hinder its larger implementation. Some of the Privacy-Preserving Artificial Intelligence (PPAI) methods including HE make it possible to secure data during the calculation process. Yet conventional HE techniques experience certain disadvantages at present with applicability to highly scalable and resource-limited applications. Moreover, this paper presents an HHE technique that is designed by integrating symmetric cryptography with HE to overcome the above-mentioned challenges successfully. To this end, we propose the GuardAI framework for end devices with limited resources such that encrypted data can be classified while preserving the privacy of input data and AI models. To show the effectiveness of the HHE, we apply it to the actual problem of heart disease classification based on the easily contaminated ECG signals. In this way, the proposed method maintains the privacy of the data with little computational and communication cost for analysts and devices and has a fairly reasonable level of accuracy in comparison with unencrypted inference. This work therefore provides a foundation for secure and private approach in AI especially for those developed to suit devices and systems with limited resources by incorporating HHE into the PPAI systems.