Lightweight Privacy-Preserving AI Using Hybrid Homomorphic Encryption: Design and Evaluation of Guard AI

Abstract

Artificial Intelligence (AI) has quickly revolutionized many industries, yet the growing susceptibility of AI models to adversarial attacks and privacy breaches is a significant obstacle to its proliferation. Homomorphic Encryption (HE) and other Privacy-Preserving Artificial Intelligence (PPAI) algorithms, which can compute on encrypted data, facilitate the implementation of privacy-preserving algorithms, but their implementation and application are typically limited by high computational and scaling costs, particularly on resource-limited systems. To overcome these drawbacks, this paper proposes a lightweight privacy-preserving AI model, Guard AI, on a Hybrid Homomorphic Encryption (HHE) scheme, which integrates both symmetric and homomorphic operations. The proposed design will minimize the computational complexity but ensure high data confidentiality when inferring the model. Guard AI is specially designed to support edge and low-resource devices and can classify data safely on encrypted data with no exposure of sensitive inputs or model parameters. To test the proposed framework, the given framework is applied to a practical healthcare application of health care use case on heart disease classification based on electrocardiogram (ECG) signals, which is very sensitive and highly vulnerable to privacy breaches. Through experiments, it is proven that the proposed HHE-based solution is characterized by a good trade-off between security, efficiency, and accuracy, where communication and computation overhead are low in comparison with other, HE schemes, whilst being competitive in comparison with inference without encryption. On the whole, this paper offers a scaled and effective framework of implementing privacy-sensitive AI systems in resource-constrained settings, indicating the promise of hybrid encryption methods in facilitating secure and lightweight intelligent systems.