Belfort Releases Fastest Encrypted Image Classifer (belfortlabs.com)

Belfort has announced the release of the "so far" CIFAR demo, an encrypted implementation of the ResNet-20 image classification model, which boasts impressive performance—outpacing the current state-of-the-art by three times with a latency of just 163ms. This breakthrough marks a pivotal moment for the AI/ML community, akin to the impact of AlexNet in 2012, signaling that encrypted AI has transitioned from a theoretical concept to a practically deployable solution. The accuracy of the model remains high, with a performance rate of 92.5%, closely matching the non-encrypted variant. The new classifier is built on Belfort’s upcoming GPU library, Cyclops, designed specifically to optimize encrypted AI workloads. As CTO Michiel Van Beirendonck notes, this advancement signifies just the beginning of a journey towards making encrypted computing feasible for larger models and broader applications. Collaborations, such as with Google's HEIR compiler, aim to streamline the use of encrypted computing, enhancing its accessibility for machine learning practitioners. Organizations interested in early access to this technology can connect with Belfort as they continue to evolve their offerings.