Researchers have proposed a bandwidth limit to protect AI from leaks

Researchers presented a methodology for constraining physical side channels, aimed at enhancing security during AI system verification. A hard bandwidth threshold of 10^5 bits/s per computational node (pod) is proposed as a defensive barrier.

During the analysis, the authors assessed the potential of various physical channels through which unauthorized information transmission could occur. It was found that optical channels, power lines, and data storage systems, in the worst-case exploitation scenarios, significantly exceed the established limit of 100 kbits/s. This creates serious risks for ensuring model isolation during their training and operation.

The research results show that current hardware architectures do not provide a sufficient level of protection against side-channel attacks. Exceeding the established threshold allows attackers to extract critical data using physical parameters of equipment operation, such as power consumption fluctuations or electromagnetic radiation.

Implementing the proposed constraint requires a revision of data center and hardware accelerator design standards. Establishing a limit of 10^5 bits/s will allow verification system developers to control data flows that cannot be tracked by software methods, ensuring more reliable isolation of computational environments from external influences.