Ix is a secure, scalable machine learning framework that creates unified information from siloed data—enabling members to leverage and contribute to collective intelligence while keeping their data private within their existing infrastructure.
Left) Original data. Right) Ix simulated output by inference step. Ix recovers the joint distribution of the image by combining two models that never see both sides of the data—model X sees only the pixel x-coordinate, and model Y sees only the y-coordinate.
Bypass the security risks and legal/statutory risks associated with sharing data, and learn across organizations, sectors, and geographies without exposing sensitive data. Create unified explainable models that allow you to learn what has been, until now, unknowable.
Keep your data where it belongs—in your secure systems. Ix requires temporary encryption only for foreign keys, which are used to create handshakes between models. All other data are secure without encryption.
Ix identifies and attributes errors and anomalies, and monitors your data in real time. Free up time to do the fun stuff while reducing exposure to incidents and litigation.
Install Ix on your current infrastructure. No special hardware like GPUs, TPUs, or FHE accelerators required.
Ix learns a holistic model of all attached data systems. This model extracts the information within your data systems, helps you continuously identify and monitor erroneous and anomalous data, and de-risk opportunities for automation and predictive analytics within your org.
Connect to the Ix-global machine learning mesh to pull in external information for better local model performance, and for access to queries based on external features.
*Compared to zama concrete-ml model trained using FHE.
Today cross-organizational inference involves coordinating across orgs with similar data features. Orgs collectively decide which prediction(s) they want to model, then they run federated learning (FL) on their encrypted data using homomorphic encryption (HE) by remotely averaging black box models during inference. In addition to adding extreme computational overhead, FL+HE is inflexible and uninterpretable.
| Feature | Ix | FL+HE |
|---|---|---|
| Distributed computation | Yes | Yes |
| Allows disjoint features to be distributed | Yes | No |
| Hides operational intent | Yes | No |
| Handles tera/petabyte data | Yes | No |
| Answer many questions with one model | Yes | No |
| Exact inference | Yes | No |
| Intepretable model with uncertainty quantification | Yes | No |
Clinical and health AI systems often fail due to the impossibility of obtaining high-quality, diverse, and representative datasets that account for real-world variations in patient populations, treatment protocols, and healthcare practices across different regions. This leads to algorithms that make inappropriate or unsafe recommendations when deployed. Ix enables secure access to all connected health information, mitigating these issues.
By connecting to cross-sector information, Ix brings in relevant external information derived from external databases linking an individual's diet, residence, travel, work schedule and history, and more to their health records, in turn reducing confounds that can cause clinical systems to break down—all without revealing any individual data.
Ix enables easy secure cross-org information sharing, maximizing threat detection likelihood. The only thing participating orgs need to know about collaborators' data stores is the presence (not content) of common indexable feature. For example, if Org A has a Personnel database with phone numbers and Org B has a Comms database with phone numbers, Org A and Org B can share information without knowing anything about each other's data.
Ix is built for real-time error and anomaly detection. Using metalearning, Ix can categorize types of anomalies to help identify anomalies specifically resulting from malicious activity.
It's not easy to swap out hardware in highly secure systems. Ix is designed to be fast on any 64-bit CPU architecture: laptops, desktops, servers, phones, smart toilets—whatever.
Baxter is a US Navy veteran and holds a PhD in Experimental Psychology and completed Postdoctoral research in AI/ML at MIT and Rutgers. He has led a number of DARPA projects and brings 15 years of experience deploying human-inspired AI/ML tech in high-risk industries.
Patrick is a program manager at DARPA under the Information Innovation office (I20) and professor of Data Science at Rutgers University - Newark. His publications have appeared in top journals of machine and human learning.
Michael has 14 years of research and engineering experience. He has built production ML systems for healthcare, agronomy, finance, and law; and has conducted research in the areas of physics, differential geometry, and high-performance computing.
Ix is currently building invite-only exchanges in health, finance, and NATSEC. Please provide your information below to receive company updates and notifications about beta availability.
