Google is making private AI practical with homomorphic encryption
Points and comments are a snapshot, not live.
Google's open-source HEIR compiler enables practical private AI inference via homomorphic encryption.
Google released HEIR (Homomorphic Encryption Intermediate Representation), an open-source compiler that converts pre-trained AI models to operate on encrypted inputs without exposing data. Four private inference applications were demonstrated: deep learning recommendation models, credit card fraud detection, threat intrusion detection (Kitsune), and hotword detection. Latency numbers were provided for single-threaded CPU. HEIR partnerships include hardware accelerators (Belfort, Niobium, Cornami, Optalysys) and academic collaborations with Georgia Tech, CMU, UC Santa Barbara, and others. Four peer-reviewed publications were built on HEIR.
The article frames homomorphic encryption as a cryptographic solution to the trade-off between data privacy and service capability, applicable to sectors like healthcare and finance. Google positions HEIR as a step toward making encrypted inference accessible to non-experts, building on its history of privacy technologies.
What commenters are saying
Commenters split into two camps: those who see FHE as a genuine cryptographic solution eliminating the need for trust, and those skeptical of Google's motives given its advertising business model and history of trust breaches. Supporters argued FHE provably prevents servers from accessing plaintext, with use cases including biometric authentication, credential checking, and private DNA matching. Skeptics countered that even with encryption, users lose autonomy (e.g., account blocking) and questioned whether Google would resist monetizing the data flow.
Technical points included: FHE guarantees confidentiality but not availability or integrity of computation; ciphertexts can be IND-CPA secure; computation can be adversarial if inputs are out of distribution. Some commenters noted that while FHE is computationally expensive, the cost is rapidly decreasing. The thread also referenced multi-party computation and oblivious transfer as related approaches.