The widespread encryption of data while stored on disk and communicated through the network — often called "at rest" and "in transit" — are critical security measures to protect business and personal data. Now Intel and Microsoft hope to create a practical and usable implementation of a third measure — "in use" encryption — that could allow encrypted data to be processed without decryption.

More formally known as fully homomorphic encryption (FHE), this area of cryptography research has already produced algorithms and systems that can manipulate encrypted data in very specific ways — for, say, averaging or searching. When the data in unencrypted, the result is the same as if the operation had been performed on the plaintext data. Yet FHE is costly, with processing requiring up to a million times more work to perform — a calculation that may take milliseconds to perform will instead take hours, days, or weeks, says Rosario Cammarota, principal engineer at Intel Labs.

To make the economics more feasible, Intel and Microsoft have signed onto a multiyear initiative launched by the US Defense Advanced Research Projects Agency (DARPA). 

"If we want to enable homomorphic encryption to process general-purpose workloads at scale — real and meaningful homomorphic encryption — then we need to go to custom hardware," Cammarota says. "From the hardware standpoint, DARPA wants a reduction in the overhead that is more than five orders of magnitude."

That means speeding up the processing by a factor of roughly 100,000. For such a feat, Intel will create an application-specific integrated circuit (ASIC) accelerator chip to speed up computations on encrypted data, while Microsoft will create cloud services around the custom hardware, Intel stated in a March 8 announcement.

The DARPA initiative, known as the Data Protection in Virtual Environments (DPRIVE) program, funds teams of companies and research organizations to rearchitect the software, hardware, and algorithms to create a platform that dramatically speeds up the computations and makes FHE a practical encryption solution, said Dr. Tom Rondeau, DARPA program manager, in a March 2 statement. 

"DPRIVE is looking to solve a really hard technical challenge that will involve a deep understanding of mathematics, algorithms, software, hardware, and circuit design," Rondeau said. "I expect that there are very few organizations that have the needed expertise in all of these areas, which are each critical to the program's success. As a result, I anticipate very interesting teams will form to cover the breadth of the research."

Limited implementations of homomorphic encryption exist, tailored for searching or for aggregating data. In September, researchers at the Massachussetts Institute of Technology announced they had finished a small pilot of a security-data sharing technology, known as the Secure Cyber Risk Aggregation and Measurement (SCRAM) system. The system allows companies to share security data without revealing the information being shared.

FHE allows companies to exchange encrypted data, or cryptograms, that can be used for specific tasks without exposing the actual data. Solving the speed problem would allow FHE to be used for computationally intensive applications, such as creating machine-learning models using encrypted data collected from a variety of sources. 

In December, Intel talked about its research into homomorphic encryption and another technology — federated learning — that could be applied to the training of machine-learning models in the future. 

FHE uses a specialized type of encryption, known as lattice cryptography, that encodes data using complex mathematical computations that are not able to be solved by current decryption techniques. However, the latest FHE algorithms use a data representation known as Large Arithmetic Word Size (LAWS), which uses data widths of thousands of bits to help mitigate some of the challenges of the algorithms. Because the word size is much longer than the 64-bit data pipelines in current processors, the standard computing system is not suited to processing fully homomorphic encryption.

Such considerations mean that a specially made processor is necessary to significantly lop a few zeroes from the computation time, says Cammarota.

"Homomorphic cyphertext — call them cryptograms — are big and ugly," he says. "To speed up their execution, we approach the problem by looking at different layer of abstractions, providing a solution at a very high level of specialization and parallelism."

The custom ASIC will be optimized for the calculations in much the same way that floating-point units (FPUs) extended the capabilities of early computers in the 1980s. 

The Intel-Microsoft team is not the only one rushing to produce a hardware accelerator for privacy-preserving encryption. Newark, NJ-based Duality Technologies announced in February that it would lead a team to develop its own FHE accelerator, known as Trebuchet. Team members include the University of Southern California Information Sciences Institute, New York University, Carnegie Mellon University, SpiralGen, Drexel University, and TwoSix Labs.