At the 22nd IEEE International Conference on Dependable, Autonomic and Secure Computing (DASC 2024), Antonis Michalas from Tampere University, representing the HARPOCRATES project, presented research titled “A Pervasive, Efficient and Private Future: Realizing Privacy-Preserving Machine Learning Through Hybrid Homomorphic Encryption.”
The research addresses privacy risks in Machine Learning (ML), which is increasingly targeted by attacks. While Privacy-Preserving Machine Learning (PPML) methods, such as Homomorphic Encryption (HE), provide protection, their inefficiency and resource demands limit their practical use on devices with constrained resources.
To address these limitations, the paper introduces Hybrid Homomorphic Encryption (HHE), which combines symmetric cryptography with HE to improve efficiency. By offloading computationally heavy tasks to the cloud, HHE enables scalable and efficient privacy-preserving ML protocols suitable for edge devices.
Key contributions include:
- Improved Efficiency: Protocols were evaluated using a dummy dataset, demonstrating significantly reduced communication and computational costs compared to traditional HE methods.
- Real-World Validation: A Privacy-Preserving Machine Learning model was implemented to classify heart disease from sensitive ECG data, showcasing its practical application in privacy-sensitive scenarios.
This work aligns with HARPOCRATES’ mission to advance secure and privacy-preserving technologies, highlighting how HHE can make ML systems more secure and efficient.
The paper is available on the HARPOCRATES Zenodo repository: Access the Paper.