Privacy usually costs speed, which is the usual reason it gets skipped. Fully homomorphic encryption keeps data hidden during computation, at a heavy performance cost. Our Equivariant Encryption paper takes another route. 📍 Inputs, outputs and intermediate activations stay unreadable to the server running the model, and the paper reports near-zero extra latency compared with unencrypted inference. The paper also includes a threat analysis that lays out the attack vectors, so anyone can test the claims.
