Google has showcased HEIR, the Homomorphic Encryption Intermediate Representation, an open source compiler that converts pre-trained AI models to operate directly on encrypted inputs. The technology lets servers process ciphertexts and return encrypted results without ever seeing the underlying data, closing the gap between privacy and AI features like spam detection or content recommendations. Google positions it as the latest addition to its private computing toolkit, alongside differential privacy and private information retrieval.
The company first announced HEIR in 2023 and says the homomorphic encryption community has since embraced it. Partners include hardware accelerator makers Belfort, Niobium, Cornami, and Optalysys, and research collaborations span Georgia Tech, Carnegie Mellon, Purdue, and other universities. To demonstrate progress, Google compiled four private inference applications with HEIR: a deep learning recommendation model, a credit card fraud detector, an anomaly detection system for encrypted network traffic, and a privacy-preserving hotword detector.
Google's stated goal is to make HEIR a one-click solution so non-experts can add encrypted inference to production apps, without needing a team of cryptographers to convert programs by hand. The cost of homomorphic encryption is rapidly decreasing, shifting the privacy versus capability tradeoff to a question of cost rather than feasibility. The announcement is one of the most discussed AI stories on Hacker News today.