A 10.33 J/encryption Homomorphic Encryption Engine in 28nm CMOS with 4096-degree 109-bit Polynomials for Resource-Constrained IoT Clients

Homomorphic Encryption (HE) is used to protect sensitive client data during offloaded compute on a potentially untrusted server. Unfortunately, the computational intensity of HE operations quickly depletes the limited resources on IoT clients. Thus, we present an energy-efficient silicon implementation of encryption/decryption in the Brakerski-Fan Vercauteren HE scheme. To support meaningful applications, including several machine learning workloads, we optimize for fixed parameters N=4096 and log(2)q=109 through pipelining, multi-level parallelism, and efficient memory accesses. At an energy-optimal operating point of 60MHz and 0.64V our chip, fabricated in a 28nm bulk planar CMOS process with an accelerator core area of 1.69 mm(2), consumes 10.33 mu J/encryption. Ultimately, this work enables IoT clients to participate in privacy-preserving offloaded compute using client-aided Homomorphic Encryption.