Enforcing Attestable Workflows across Untrusted Networks
JANUS is a high-performance architecture that federates Trusted Execution Environments across untrusted networks using a split Trusted Computing Base design, achieving native-speed encrypted routing with minimal latency overhead by anchoring key release to hardware measurements and executing enforcement within the kernel.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern digital world, some of the most valuable work happens in the cloud. Scientists analyzing medical images, researchers training artificial intelligence, and engineers simulating climate patterns often need to combine data from many different computers, sometimes owned by different organizations. To keep this work private, they rely on special hardware that creates a secure, locked room inside a computer processor. This technology, known as a trusted execution environment, ensures that even if the computer's operating system is compromised, the data inside this locked room remains hidden. However, a new problem arises when these locked rooms need to talk to each other across the internet. Traditional methods of securing this communication often force the data to be decrypted and re-encrypted at every step, slowing everything down, or they require the software to be rewritten entirely to fit a specific security model. The challenge is to keep the data locked while it travels between these secure rooms, without slowing down the massive amounts of information scientists need to move.
A team of researchers has developed a new system called Janus to solve this specific problem. The system is designed to let secure computers talk to each other at full speed while proving that they are who they say they are and that they are following the rules. The researchers built this system to work with the latest generation of secure hardware, specifically Intel's TDX technology, which creates these locked rooms for entire virtual machines. The core idea behind Janus is to split the job of security into two parts. One part handles the heavy lifting of proving identity and exchanging secret keys, while the other part handles the actual movement of data. By moving the data handling deep into the computer's core, where it can operate at the speed of the hardware itself, the system avoids the slow, clunky steps that usually happen when security software runs on top of the operating system.
To test how well this works, the researchers set up a large experiment using thirty-two powerful computers in the cloud. They created a simulated environment where these computers needed to share sensitive data, much like a team of scientists working on a shared project. The system worked by first having the computers prove their identity to each other using a special handshake. This handshake involved the hardware generating a unique, unforgeable report of its current state. If the reports matched the expected rules, the computers would exchange a secret key to lock their conversation. Once this initial setup was complete, the data began to flow. The researchers found that the system could handle this encrypted traffic with almost no slowdown. For every single packet of data moving through the network, the system added a delay of only six microseconds. This is an incredibly small amount of time, meaning the computers could process data at nearly the same speed as if they were sending it without any encryption at all.
The results showed a stark difference between this new approach and older methods. When the researchers compared Janus to existing systems that rely on software running in the user space, the older systems were significantly slower. Those traditional methods added a delay of between 137 and 147 microseconds for every packet, which is more than twenty times slower than the new system. In a real-world test involving a complex machine learning task that required moving over a gigabyte of data, the new system slowed down the total process by only 6.1 percent. In contrast, the older methods slowed the same task down by 62 percent. This efficiency is crucial for scientific workflows that run for hours or days, where even a small delay can add up to a significant loss of time and computing power.
The system also proved to be very fast at getting started. In a simulated network of one hundred computers, the entire group was able to establish secure connections and begin working in less than one and a half seconds. This speed was achieved by running the identity checks in parallel across the different machines rather than waiting for them to happen one by one. The researchers also tested how the system handled mistakes or changes in the network. If a computer tried to connect with the wrong software or if the security rules changed, the system immediately blocked the connection without crashing or leaking data. It was designed to be resilient, meaning that if a computer failed or the network had a temporary glitch, the rest of the system would keep working, and the failed connection could be re-established quickly once the issue was resolved.
One of the most important aspects of the research is what the system does not do. The researchers explicitly showed that their approach does not require scientists to rewrite their software programs. The system works transparently, meaning the applications running on the computers do not need to know that they are being secured. The security happens underneath, in the network layer, invisible to the user. This is a significant departure from other security frameworks that force developers to change their code to fit a specific security model. By keeping the applications unchanged, the system allows existing scientific tools to be used in secure environments without modification.
The study also highlighted the limits of the current technology. While the system works very well for networks where computers connect to a few others, the researchers found that if every computer tried to connect to every other computer at the same time, the initial setup would slow down significantly. This is because the hardware used to generate the identity reports has a limit on how many it can process at once. For the types of large-scale scientific pipelines the researchers had in mind, where data flows in a specific path rather than a chaotic web, this limitation was not a problem. The system is optimized for these long-running, structured workflows, ensuring that the security does not become a bottleneck for the science.
Ultimately, the work demonstrates that it is possible to build a secure, high-speed network for confidential computing that does not sacrifice performance for safety. The researchers showed that by carefully designing how the security checks and the data movement interact, they could achieve a balance that was previously thought difficult to reach. The system provides a way for organizations to share sensitive data across untrusted networks with the confidence that the data remains encrypted and the participants are verified, all while maintaining the speed required for modern high-performance computing. This opens the door for more collaborative scientific research, where institutions can work together on sensitive problems without compromising the privacy of their data or the efficiency of their work.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.