"Machine Learning Helps Fuzzing Find Hardware Bugs"

A new study titled "MABFuzz: Multi-Armed Bandit Algorithms for Fuzzing Processors" delves into the fuzzing technique that introduces commands and prompts to a chip in order to cause the system to behave erratically and unpredictably. The irregular responses to "fuzzed" commands can then point researchers to potential vulnerabilities in the system. If the system does something unusual, researchers examine it to see if there is a security flaw that hackers could exploit. Researchers used reinforcement learning to select inputs for fuzz testing in this study. They adapted an algorithm that is used to solve the Multi-Armed Bandit (MAB) problem, which is the difficulty of optimizing rewards when given the option of accepting known rewards or exploring rewards that may be greater or lower. This study uses the MABFuzz algorithm to decide whether to try a new random input or stick with an input that works well. MABFuzz was found to significantly accelerate the detection of vulnerabilities. This article continues to discuss key findings and observations from the study.

IEEE Spectrum reports "Machine Learning Helps Fuzzing Find Hardware Bugs"

Submitted by grigby1