Google Advances Private AI Inference With HEIR Compiler

Homomorphic encryption allows computation on data while it remains encrypted, reducing the risk that cloud providers can view sensitive personal or business information. Adoption has been constrained by steep performance costs and the need for specialists to rewrite software around complex cryptographic schemes. Google’s open-source HEIR compiler project seeks to automate much of that work, creating a pathway for existing programs and artificial intelligence models to operate on encrypted data without exposing the underlying inputs.

Google disclosed new HEIR capabilities that can convert existing pretrained AI models into versions designed to process encrypted inputs directly, without requiring a cloud service to decrypt the original data. The company highlighted potential uses including content recommendations, credit-card fraud detection and speech recognition. By shifting model conversion and cryptographic optimization into the compiler toolchain, HEIR could lower the engineering barrier to privacy-preserving inference, though Google did not disclose commercial deployment dates, performance benchmarks or adoption figures in the information provided.