Efficient performance estimation on fully homomorphic encryption accelerators for optimal parameter setting

Abstract

As the size of ML models grows and the server handles nearly all computations, research on the user privacy has become crucial. Fully Homomorphic Encryption (FHE), enabling operations on encrypted data and providing quantum-safe security, is considered the most promising methodology. However, operations with FHE are significantly slower—ranging from 10,000 to 100,000 times slower than plaintext operations—prompting the research on FHE hardware accelerators. The performance of accelerators is highly influenced by the parameter settings for the ciphertext space. However, the parameter settings explored in existing accelerators are optimized only for specific security levels, often falling short of achieving optimal performance. Additionally, these parameters are intricately related, making design space exploration challenging and involving a vast number of possible scenarios. This thesis proposes a framework to efficiently predict the performance of an accelerator under given conditions and, further, to discover the optimal parameter settings. The implemented framework first identifies and abstracts key elements determining the accelerator’s performance. Subsequently, it simulates and predicts the performance of the accelerator for specific parameter settings. With the proposed framework, this thesis demonstrates the ability to obtain optimal parameter settings based on various security requirements not previously explored in existing accelerators.