The dream of running a neural network on encrypted data — a hospital sending an encrypted scan to a cloud model and getting back an encrypted diagnosis the cloud never reads — keeps colliding with the same wall: speed. Homomorphic operations are orders of magnitude slower than their plaintext equivalents, and a neural network is millions of multiplications. Feasible in principle has long meant unusable in practice.
US11405176B2, “Homomorphic encryption for machine learning and neural networks using high-throughput CRT evaluation,” granted to Intel Corporation on August 2, 2022, goes after that wall with a classic technique. Classified under H04L 9/008, it claims using the Chinese Remainder Theorem to make the underlying arithmetic high-throughput.
The Chinese Remainder Theorem is centuries-old number theory, and its role here is to divide and conquer. The giant numbers homomorphic schemes operate on can be represented as several smaller residues modulo coprime values; arithmetic can be done independently on each small piece and recombined. On hardware with many parallel lanes — exactly what Intel builds — that decomposition turns one slow big-integer operation into many fast small ones running at once.
Reading the claim against the title, the load-bearing word is “throughput.” This is not a new homomorphic scheme; it is a way to make an existing one run fast enough on real hardware to matter. The inventor team — Ghosh, Reinders, Misoczki, Cammarota, Sastry — overlaps with Intel's other cryptographic-acceleration filings, reinforcing that this is about performance engineering, not new math.
Per the desk's discipline: issued grant (B2), not an application; method claim, not a shipped product. Intel's homomorphic-encryption acceleration research and its HEXL library are the plausible context, but the patent is the technique.
The strategic read is that the homomorphic-ML race, like the rest of the field, is being decided on throughput. Whoever owns the efficient hardware-friendly evaluation methods owns the practical version of the technology. A CRT-based throughput patent from a chipmaker is precisely the sort of IP that decides whether encrypted machine learning ever leaves the lab.