Google is making private AI practical with homomorphic encryption

Google has unveiled HEIR, a powerful open-source tool added to its Private Computing Toolkit, enabling secure and private AI inference. As AI grows, ensuring privacy and security becomes crucial, and standard protections like end-to-end encryption have limitations. HEIR addresses this by using homomorphic encryption, which allows computations on encrypted data without exposing the information.

This technology shifts the privacy and capability trade-off to a cost question, and its costs are decreasing rapidly. Google's history of privacy technology innovations, such as differential privacy and secure enclaves on Google Cloud, led to the development of HEIR. HEIR compiles pre-trained AI models to operate on encrypted inputs, making it accessible to non-experts.

Google has partnered with companies developing hardware accelerators for homomorphic encryption, and the tool has been adopted by various research institutions. Four private inference applications compiled with HEIR demonstrate its effectiveness, with latency numbers provided for a single-threaded CPU. The source code for these examples is available on GitHub, showcasing the potential of homomorphic encryption to become ubiquitous in the industry.