Today, we are thrilled to introduce HEIR, an innovative addition to our suite of Private Computing tools. HEIR is an open-source compiler that facilitates cryptographically-secure private AI inference.
As artificial intelligence continues to advance, the critical balance between privacy and security has become increasingly prominent. Traditional methods like end-to-end encryption, while effective at safeguarding user data from breaches, often limit the capabilities of service providers who rely on that data for functionalities like spam detection or virus protection. This challenge is even more pronounced in sensitive sectors such as healthcare and finance, where stringent regulations restrict data sharing between organizations. Alternative solutions such as local data processing often fall short due to device limitations and the need to protect proprietary information, which poses a risk of exposing AI models when deployed on local devices.
Homomorphic encryption emerges as a transformative technology that redefines this balance by enabling computations on encrypted data. With this method, servers can process encrypted inputs and deliver results without revealing any underlying data. For instance, a cloud service can suggest content based solely on user input without ever accessing that information. One of our demonstrations highlights this capability precisely. Although the implementation of homomorphic encryption usually involves some cost overhead, it effectively turns the privacy versus capability debate into a matter of cost, which is steadily decreasing.
Google has a longstanding tradition of pioneering privacy-enhancing technologies, with innovations ranging from differential privacy to secure enclaves on Google Cloud. Homomorphic encryption adds another layer to our Private Computing Toolkit. This method, unlike hardware-focused techniques, offers robust security and privacy assurances rooted in cryptographic principles. However, translating an existing application to utilize homomorphic encryption effectively necessitates an adept team of cryptographers.
To address usability barriers and harness the potential of homomorphic encryption, our talented team at Google created the HEIR compiler project. HEIR (Homomorphic Encryption Intermediate Representation) is an open-source toolchain and development platform designed to support homomorphic encryption. Specifically, HEIR enables the conversion of pre-trained AI models designed for unencrypted data to work with encrypted inputs. Our goal is to provide a straightforward, one-click solution that empowers users without technical expertise to integrate encrypted inference into their applications seamlessly.
Since its announcement in 2023, HEIR has garnered support from the homomorphic encryption community. We have collaborated with companies specializing in hardware accelerators for this technology, including Belfort, Niobium, Cornami, and Optalysys. These partnerships have yielded impressive demonstrations, and we anticipate showcasing the latency advantages of these accelerators soon. Additionally, HEIR has proven to be a valuable research platform, allowing cryptographers to focus on optimization while utilizing existing infrastructures for testing and benchmarking. This collaborative effort has allowed us to work with institutions such as Georgia Tech, Carnegie Mellon, UC Santa Barbara, the Illinois Institute of Technology, Purdue, and the University of Edinburgh, among others. To date, four peer-reviewed articles have been published based on HEIR, with more in the pipeline.
In demonstrating the advancements in homomorphic encryption, we’re excited to present four private inference applications, all compiled using HEIR. Each application highlights the efficiency and potential of this technology, with latency metrics available for single-threaded CPUs. The source code for these projects can be found in our GitHub repository.
1. A Deep Learning Recommendation Model enables personalized content recommendations through collaboration with Belfort Labs, LG, and New York University. 2. A credit card fraud detection system was developed in partnership with Niobium and hardshell.ai. 3. The Kitsune system for anomaly detection in encrypted network traffic, also created with Niobium, allows service providers to identify irregularities without accessing the actual data packets. 4. A hotword detection model compiled in collaboration with Belfort Labs, which helps audio-triggered AI agents recognize specific commands while maintaining the privacy of the audio feeds.
As the software industry evolves amidst increasing demands for security and privacy in the age of AI, our research team is dedicated to making homomorphic encryption more accessible, efficient, and widely adopted across various sectors.
