Google just made an announcement that has been part of the long march of encrypted computing maturation and adoption. Their new Google HEIR compiler converts pre-trained AI models to run on encrypted data, greatly reducing the technical effort need, with demos built alongside hardware accelerator companies like Niobium, Belfortt, Cornami, Inc., and Optalysys. Google's framing: homomorphic encryption is "rapidly maturing," and the trade-off is now cost, not capability. And the cost is falling fast. When we started this work under DARPA programs that invested in Fully Homomorphic Encryption (#FHE) in the early 2010s, FHE was originally a mathematical curiosity, millions of times too slow. Over a decade, a layered ecosystem emerged: open-source libraries (including OpenFHE, which I co-founded and which compilers like HEIR target as a backend), then hardware acceleration (Defense Advanced Research Projects Agency (DARPA)'s DPRIVE program seeded much of today's accelerator landscape). IARPA made some fundamental early investments in compilers for FHE, but the field has moved forward and now compilers like HEIR let a much wider range of developers use it all without a cryptography team. Duality Technologies has been deploying this operational FHE-protected secure data collaboration software stacks for governments and regulated enterprises for years. We're happy to support and be a part of the open-source ecosystem for FHE, along with Google and the many other collaborators and contributors in the domain. Alon Kaufman Rina S. shafi goldwasser Vinod Vaikuntanathan Yuriy Polyakov Andreea Alexandru Ahmad Al Badawi, PhD https://lnkd.in/ggz8CtDh
Google HEIR Compiler Advances Encrypted Computing with Reduced Technical Effort
This title was summarized by AI from the post below.
More Relevant Posts
-
Google open-sourced HEIR, a compiler toolchain for homomorphic encryption. It converts pre-trained models to run on encrypted data, so a server processes the ciphertext and returns an encrypted result without ever seeing the input. The overhead is still nontrivial, and Google says it is falling fast. Hardware partners include Belfort, Niobium, Cornami and Optalysys. The stated goal is one-click encrypted inference for non-experts. My thought: The next generation of conversation based solutions will need to start obfuscating personal user data before sending to models/tracing. read it → https://lnkd.in/emzvNT69
-
My friend Bryant is doing some **really** interesting research at Google on privacy/encryption that may help make some AI applications more secure. Check it out!
Locking down user data shouldn't mean locking out the features that make AI helpful. 🔐 Several years ago, we set forth a vision (Unlocking the Potential of Fully Homomorphic Encryption): FHE didn't have to remain strictly theoretical. We argued the bridge to widespread adoption was a usability gap—one we could close through unified compilation. Today, that vision is a reality. We just published our latest milestone detailing how we are making private AI practical with the release of HEIR (Homomorphic Encryption Intermediate Representation). It's an open-source compiler toolchain that allows developers to run pre-trained AI models directly on encrypted inputs—completely protecting both the model and the underlying user data. By shifting the privacy-vs-capability trade-off into a solvable hardware optimization question, we've successfully compiled four real-world private inference applications: - Private Content Recommendations (joint work with Belfort Labs, LG, and NYU) - Encrypted Credit Card Fraud Detection (collaborating with Niobium and hardshell.ai) - Encrypted Network Threat Intrusion Detection (with Niobium) - Private Hotword Detection (with Belfort Labs) Taking this from a research paper to an active open-source ecosystem has been a massive effort. A huge thank you to all of our academic and hardware partners who are building the infrastructure to make FHE ubiquitous with us. Check out the full blog post to see how we're making private AI a production reality! #FullyHomomorphicEncryption #PrivateAI #Cryptography #PrivacyPreserving #GoogleResearch #HEIR https://lnkd.in/gdmebkUS
-
Excited to be collaborating with the Google and Belfort teams to help advance Private AI. One of the biggest challenges in AI today is balancing powerful intelligence with strong privacy guarantees. If we can solve that challenge, it has the potential to unlock entirely new classes of applications that can safely leverage sensitive data. Google's HEIR project is an important step in that direction. By providing an open-source compiler that enables AI models to run directly on encrypted data, HEIR is making privacy-preserving AI significantly more accessible to researchers and developers. As part of this effort, we demonstrated an encrypted recommendation model developed through a collaboration between LG and New York University, showing how personalized content recommendations can be delivered while keeping user data protected. As a former Googler, it's especially meaningful to collaborate on work like this. Google has long been a leader in open research and ecosystem building, and I continue to believe it is uniquely positioned to help drive technologies that create broad societal impact. Looking forward to continuing the journey with great collaborators and pushing the boundaries of what's possible with Private AI. #AI #PrivacyAI #FHE #HomomorphicEncryption #MachineLearning #Google #OpenSource #Research #Innovation
Locking down user data shouldn't mean locking out the features that make AI helpful. 🔐 Several years ago, we set forth a vision (Unlocking the Potential of Fully Homomorphic Encryption): FHE didn't have to remain strictly theoretical. We argued the bridge to widespread adoption was a usability gap—one we could close through unified compilation. Today, that vision is a reality. We just published our latest milestone detailing how we are making private AI practical with the release of HEIR (Homomorphic Encryption Intermediate Representation). It's an open-source compiler toolchain that allows developers to run pre-trained AI models directly on encrypted inputs—completely protecting both the model and the underlying user data. By shifting the privacy-vs-capability trade-off into a solvable hardware optimization question, we've successfully compiled four real-world private inference applications: - Private Content Recommendations (joint work with Belfort Labs, LG, and NYU) - Encrypted Credit Card Fraud Detection (collaborating with Niobium and hardshell.ai) - Encrypted Network Threat Intrusion Detection (with Niobium) - Private Hotword Detection (with Belfort Labs) Taking this from a research paper to an active open-source ecosystem has been a massive effort. A huge thank you to all of our academic and hardware partners who are building the infrastructure to make FHE ubiquitous with us. Check out the full blog post to see how we're making private AI a production reality! #FullyHomomorphicEncryption #PrivateAI #Cryptography #PrivacyPreserving #GoogleResearch #HEIR https://lnkd.in/gdmebkUS
-
Locking down user data shouldn't mean locking out the features that make AI helpful. 🔐 Several years ago, we set forth a vision (Unlocking the Potential of Fully Homomorphic Encryption): FHE didn't have to remain strictly theoretical. We argued the bridge to widespread adoption was a usability gap—one we could close through unified compilation. Today, that vision is a reality. We just published our latest milestone detailing how we are making private AI practical with the release of HEIR (Homomorphic Encryption Intermediate Representation). It's an open-source compiler toolchain that allows developers to run pre-trained AI models directly on encrypted inputs—completely protecting both the model and the underlying user data. By shifting the privacy-vs-capability trade-off into a solvable hardware optimization question, we've successfully compiled four real-world private inference applications: - Private Content Recommendations (joint work with Belfort Labs, LG, and NYU) - Encrypted Credit Card Fraud Detection (collaborating with Niobium and hardshell.ai) - Encrypted Network Threat Intrusion Detection (with Niobium) - Private Hotword Detection (with Belfort Labs) Taking this from a research paper to an active open-source ecosystem has been a massive effort. A huge thank you to all of our academic and hardware partners who are building the infrastructure to make FHE ubiquitous with us. Check out the full blog post to see how we're making private AI a production reality! #FullyHomomorphicEncryption #PrivateAI #Cryptography #PrivacyPreserving #GoogleResearch #HEIR https://lnkd.in/gdmebkUS
-
Google has introduced the HEIR (Homomorphic Encryption Intermediate Representation) compiler project. Homomorphic encryption allows operations on encrypted data without having to decrypt/re-encrypt it. HEIR can convert pre-trained AI models that operate on unencrypted data to operate on encrypted inputs, allowing AI agents to process sensitive data securely. #cybersecuriting #encryption #homomorphicencryption #Google #HEIR https://lnkd.in/geEBxUWz
-
Google releases open source a compiler for running AI on encrypted data HEIR converts pre-trained AI models to process fully encrypted input, allowing servers to run inferences without ever seeing users' plaintext data. The tool includes four demos—including fraud detection and content recommendation—developed in partnership with Niobium, Belfort Labs, and NYU. Homomorphic encryption is still roughly 100 to 1,000 times slower than plaintext processing, and no Google consumer product currently uses HEIR. https://lnkd.in/d86R5wWf
-
Exciting to see our work HE-LRM featured as a use case in Google’s latest Security Blog post on making private AI practical with Homomorphic Encryption! https://lnkd.in/giXEgwd9 In collaboration with Belfort and New York University (Prof Brandon Reagen, Karthik Garimella and Austin Ebel), LG Electronics North America ETL Advanced Security Team explored privacy-preserving recommendation, enabling recommendations to be computed directly on encrypted user data without exposing the underlying information. Happy to see more and more homomorphic encryption use-cases moving from research to practical, real-world AI applications.
-
Google is making private AI practical with homomorphic encryption "Today we're excited to showcase HEIR, the latest powerful tool added to our Private Computing Toolkit. HEIR is an open source compiler that unlocks cryptographically-secure private AI inference." https://lnkd.in/gvMUk47C
-
This is one of the coolest security innovations I've read about in a long time: https://lnkd.in/gmKn9raj If they're able to reduce things like cost and latency to an acceptable level, this could mean a massive set of security and privacy problems will be solved, ultimately making user-facing apps much more safe. Performing inference on encrypted data (meaning the AI provider has no idea what the inputs are) is insanely awesome for privacy and security purposes. I wish there were easier ways to deploy homomorphic encryption for standard user systems as easily as this -- imagine a world where all the Google Docs you build, all the slides you create, and all the photos/videos you capture can still be edited with best-in-class tools without the provider ever "seeing" the raw contents. That's pretty amazing stuff.
