Quantifying the Bit-Error Resilience of FHE Compute

Homomorphic Encryption (HE) enables computation on encrypted data paving the way for a solution for privacy-preserving computation, especially in non-trustable environments. As HE gradually makes its way through real-world applications and problems, like machine learning and AI, ensuring the robustness and resilience of the technique becomes critical.

In this talk, we present a study and characterization of the sensitivity of HE to bit errors across the encoding, encryption, decryption, and decoding stages. Preliminary results and experimental observations provide some insights on the fundamental aspects of HE robustness and fault tolerance and unveils opportunities to mitigate undesirable behavior in the presence of bit errors. We explain how certain HE schemes, like CKKS, exhibit different levels of resilience across different configuration parameters and scenarios. In particular, widely adopted optimizations, like the Residue Numeral System (RNS) and the Number Theoretic Transform (NTT), can severely exacerbate the error sensitivity of HE schemes.