Intel HERACLES
Rosario Cammarota
Abstract
In spite strong advances in confidential computing technologies to protect data, the status is that data is encrypted while temporarily not in use and unencrypted during computation. The inability to keep data encrypted during computation can inhibit the ability to fully share and extract the maximum value out of data, e.g., via statistical and machine learning methods with the additional risk of third-party data leakage. Fully Homomorphic Encryption (FHE) enables users to delegate computation to the cloud by enabling the cloud to process users' encrypted inputs without decryption and return encrypted output to the intended recipients. The adoption of FHE by the industry has been slow. First, processing encrypted data incurs a huge performance tax even for simple operations - ciphertexts operations can be several orders of magnitude slower with respect to cleartext operations on existing hardware. Second, there is lack of automation tools for translating data and applications to enable FHE. Third, there is lack of standards and best practices for FHE secure deployment in combination with other confidential computing techniques. The DARPA DPRIVE program [1] is the first publicly visible program aiming to build a platform to accelerate FHE and a path to their commercialization, for their use in healthcare, communication (5G to XG), and cloud computing. The program represents a steppingstone toward FHE adoption. Under the DARPA DPRIVE program, Intel® is designing a new type of computer architecture to reduce the performance tax currently associated with FHE. Intel® collaborates with Microsoft® Azure Global in realization of the project [2]. The design includes flexible arithmetic circuits for algebraic lattices with unprecedented vector parallelism and data transfer capacity between vector slots, to increase ciphertext processing speed, coupled with near-memory computation, to reduce data movement. The software stack leverages the Microsoft® SEAL library [3] augmented the BGV mechanism including bootstrapping [4], and automatic translation tools to explore trade-offs in algorithmic optimization and data encoding to fit the performance requirements [5]. When fully realized, the HERACLES platform will deliver a massive improvement in executing FHE workloads over existing CPU-driven systems, potentially reducing ciphertext processing time by five orders of magnitude. Beyond the new hardware and software stack, Intel® work with international standard bodies to develop standards for and best practices for FHE secure deployment [6] and invests in academic research to advance both theory and application [7].