Niobium brings fully encrypted AI workloads to the cloud with The Fog – News

A startup called Niobium Microsystems Inc. is bringing “The Fog” to the cloud, enabling organizations to run artificial intelligence and data processing workloads on their most sensitive data without ever needing to decrypt it.

Though the new platform might sound like a 1980s-era horror flick, Niobium says The Fog is actually something developers will want to embrace. Slated to launch later this year, it’s the first platform of its kind to leverage fully homomorphic encryption or FHE. Unlike traditional encryption methods that can only protect data in transit or at rest, FHE makes it possible to encrypt data as it’s being processed inside a server or virtual machine, ensuring that even the most sensitive information remains private at all times.

Niobium says The Fog provides a “zero trust” advantage, because the decryption keys will always remain with the owner of the data, ensuring that no third parties, not even itself, can see the information being processed.

The startup aims to solve a major challenge for organizations, which are sitting on massive volumes of sensitive data that cannot easily be tapped for insights. To analyze sensitive data, companies normally have to decrypt it first, but that means exposing it to potential security breaches and insider threats, and going against compliance regulations. It’s a particular challenge in highly regulated industries such as healthcare and financial services, making them hesitant to move their most sensitive workloads to cloud environments.

Chief Executive Kevin Yoder said many companies just accept the risk of data exposure as the cost of doing business in the cloud. “The Fog eliminates that tradeoff,” he promised. “Our goal is to make encrypted computing practical, scalable and accessible to the teams that need it most.”

Democratizing FHE

Though FHE is not a new innovation, the technique has traditionally involved massive computational overheads that make it too slow and prohibitively expensive to use at scale. Niobium’s core value proposition is that it overcomes this barrier. It has developed a new processor called the “mistic Core,” which is based on a field-programmable gate array architecture. Available in private beta now, the chip enables FHA tasks to run twice as fast as any graphics processing unit or specialized accelerator.

“Live data encryption has traditionally always been very expensive and cumbersome, slowing down applications,” said Holger Mueller of Constellation Research. “For a long time, FHE was the Achille’s Heel of data encryption because of this, and so it’s good to see the innovation with the fittingly named The Fog offering. The next question is how quickly enterprises will adopt this technique.”

To help enterprises get started with FHE, Niobium has created a number of template applications. These include an Encrypted Semantic Search app that makes it possible to query sensitive data by meaning instead of exact match. It ensures both the query and the underlying data remain fully encrypted, making it ideal for secure retrieval-augmented generation workloads.

There’s also a Federated Learning app that makes it possible to train AI models on distributed datasets without ever exposing them. Finally, the company has developed a Machine Learning Classification app that can analyze encrypted information to identify patterns, trends and security threats.

The launch of The Fog reveals Niobium’s aspirations to transform FHE from a mathematical curiosity into an industry-standard model that will pave the way for organizations to do more with their most sensitive data. To that end, the company has launched a dedicated compiler and software development kit to help developers with no knowledge of encryption build apps that can run on the platform.

The Fog’s capabilities are especially intriguing for AI developers. AI systems are only ever as good as the data that powers them, but a huge volume of the richest datasets remain off-limits for AI models due to security concerns. Niobium opens up the possibility for companies to train shared AI models on massive volumes of pooled, highly sensitive data, because it can ensure that no party sees the actual datasets. It can also solve the question of prompt privacy, where workers sometimes upload sensitive data alongside their prompt so that a large language model can analyze it and generate a response.

In the future, Niobium says, it will make FHE even more powerful. Today it revealed it’s working with the custom silicon developer Semifive U.S. Inc. and Samsung Foundry to create a new customized application-specific integrated circuit for FHE. The startup says this will dramatically improve FHE performance, because of the way ASICs’ circuits can be optimized to run individual applications.

Images: Niobium