Accelerating Fully Homomorphic Encryption with Processing in Memory
TimeWednesday, December 8th11:00am - 11:30am PST
Special Session (Research Track)
Hosted in Virtual Platform
DescriptionFully homomorphic-encryption(FHE) provides a promising solution for future computing needs by allowing privacy-preserving computation. However,many FHE schemes work with strict security parameters providing little to no flexibility and have large performance overhead. Micciancio-et-al recently presented FHEW and TFHE frameworks for third-generation FHE that work for a range of security parameters. The output of the framework has higher noise tolerance than the state-of-the-art implementations, which is critical to ensure safe decryption. However, its practical use has been limited by the huge latency overhead it incurs while computing. Our recent work showed that PIM can accelerate FHE operations by over 900x as compared to state-of-the-art CPU implementation. Moreover, as the size of polynomials increases from 256 to 32k, CPU throughput dropped from 30000 to 100 ops. On the contrary, PIM provided immense scalability by maintaining the same order of magnitude throughput, which could be improved further by making the PIM vectors longer. As a result, we accelerate the third generation FHE using processing-in-memory(PIM). Our FHE-PIM platform assumes a client-server design and accelerates the end-to-end FHE computations on both client and server. FHE-PIM has two main components: i)the client PIM that can encode, encrypt, decrypt, and decode data given some security parameters and ii)the server PIM that can homomorphically compute on the encrypted data. It is designed to provide flexible and accelerated bootstrapping, which enables the user to employ either a simple or a complex bootstrapping procedure based on the specific application parameters. Our FHE-PIM design enables efficient computation of arbitrary functions homomorphically.