Oak Ridge National Laboratory’s Center for Artificial Intelligence Security Research, or CAISER, has introduced Photon, a framework designed to rapidly identify vulnerabilities in AI models at exascale.
How Does Photon Work?
Photon was developed by reimagining ORNL's DeepHyper technology, a tool originally used for training large neural networks. DeepHyper was repurposed to identify attack parameters and expose potential threats.
Photon applies known attacks from the published literature against target models, then refines them by uncovering new weaknesses. This cycle goes on until performance degradations are no longer observed.
How Does Photon Enhance Vulnerability Testing?
The framework uses an asynchronous, decentralized approach that allows multiple attack scenarios to run simultaneously across systems. Instead of relying on a single coordinating process, Photon enables independent “agents” to share findings in real time, enhancing the effectiveness of ongoing tests. This method enables the system to quickly examine large hyperparameter spaces.
What Role Does Frontier Play in the Effort?
Photon leverages ORNL's Frontier exascale supercomputer to support large-scale AI vulnerability testing. The framework can execute around 60,000 jailbreak prompts per hour, reducing auxiliary tasks and delay while sustaining over 95 percent resource utilization across thousands of graphics processing units, or GPUs.
