Abstract
We present a technical evaluation of a new, disruptive cryptographic approachto data security, known as HbHAI (Hash-based Homomorphic ArtificialIntelligence). HbHAI is based on a novel class of key-dependent hash functionsthat naturally preserve most similarity properties, most AI algorithms rely on.As a main claim, HbHAI makes now possible to analyze and process data in itscryptographically secure form while using existing native AI algorithms withoutmodification, with unprecedented performances compared to existing homomorphicencryption schemes. We tested various HbHAI-protected datasets (non public preview) usingtraditional unsupervised and supervised learning techniques (clustering,classification, deep neural networks) with classical unmodified AI algorithms.This paper presents technical results from an independent analysis conductedwith those different, off-the-shelf AI algorithms. The aim was to assess thesecurity, operability and performance claims regarding HbHAI techniques. As aresults, our results confirm most these claims, with only a few minorreservations.