University of Tennessee-Chattanooga doctoral candidate develops self-encrypting AI

Joshua Tyler, who is on track to receive his third UTC degree in May, has developed the world’s first usable AI network that can learn how to encrypt itself. (Photo by Angela Foster/UTC)

Joshua Tyler, a computational engineering doctoral candidate and electrical engineering research associate at the University of Tennessee at Chattanooga (UTC), has developed the world’s first usable artificial intelligence network that can learn how to encrypt itself.

According to a UTC news release, the AI network can provide nearly unbreakable cryptography, which significantly improves the security of communications.

The announcement said he and his faculty mentor, Dr. Don Reising, a Guerry and UC Foundation associate professor of electrical engineering recently submitted their invention disclosure to the University of Tennessee Research Foundation for a provisional patent. The two also uploaded a draft of their publication to arXiv, an open-access repository for scholarly papers.  

Tyler noted that his AI network learns encrypt data “by transforming an encryption key onto the original unencrypted data.” He said the goal is to “ensure that the encrypted message is unique to the key while the original message is still recoverable on the other end.”

“This ensures that when deployed, each encryption is unique and significantly extends the network’s lifespan,” said Tyler.

According to the announcement, Tyler’s research builds on a concept initially proposed by Google, known as Adversarial Neural Cryptography. While Google demonstrated the potential for AI-driven cryptography, he said the technology faced limitations when ensuring the encryption key’s influence on the encrypted message and additional communication overhead.

“I copied over Google’s setup and trained their network on my side,” he said. “The network was encrypting the information, but we found out that there wasn’t a lot of uniqueness on the encrypted side when we were switching keys, so that makes the overall life of the network shorter. You’d only get to encrypt one message per network.”

Reising, who has worked with Tyler for more than six years, said he “basically challenged Josh to go and find a way to get this thing to generate a unique code or a unique encoded message.”

“I asked him, ‘What architecture are you using? Are you using CNN? Are you using an LSTM? What are you using?,’” Reising said in the announcement. “And he’s like, ‘No, I’m not using any of those. I made my own.’

“I said, ‘What do you mean you made your own?’” he added. “He said, ‘I made my own and it’s a deep learning network.’ That was crazy and it was pretty awesome.”

The news release said that by rethinking the structure of AI networks, Tyler developed a “novel neural network architecture” that addresses encryption challenges. The result was a network that offers nearly unbreakable encryption, which helps safeguard sensitive data.

“I changed the network architecture so that the influence of the key was still maintained through the entire structure of the network,” he said.

Tyler noted that it took several months of trying to get the new system to work with the traditional architectures.

“When it finally worked,” he said, “it was surprising because you go through two or three months of it not working—and then you try something completely new that you invented and it suddenly works. That’s where the real fun starts because then you get to make sure that it really works.”

The news release noted that the implications of Tyler’s discovery could extend across industries, including defense, health care, and finance.

“In the military, they want to do a lot of unmanned vehicles, and a really big consideration is communications,” he said, noting that new aircraft are being designed to be unmanned because it costs a lot of money to put a human pilot in the cockpit.

“You have to have life support systems, you have to have G-force and you have to pay to train that pilot—and that takes millions of dollars,” he continued. “If you have an unmanned vehicle, you can go faster and you don’t risk the pilot’s life every time they go out on a mission or do training—but that requires coordination between the operating base and the unmanned vehicle. It has to communicate a lot of information, but we don’t want that information to be gleaned by any adversary.”

The announcement said the technology could secure sensitive information transmitted between devices, which could be particularly useful in health care.

“In a hospital, you have the vitals being read in the patient’s room—and that’s being fed to the nurse’s station to where they can monitor that,” Tyler said. “There may be someone that wants to intercept that information for some reason. So, we would prevent them from being able to intercept a patient’s vitals.”

To learn more about Tyler and Reising’s work, visit blog.utc.edu.