FeedbackFuzz: Fuzzing Processors via Intricate Program Generation with Feedback Engine

Abstract

As modern processor designs become increasingly complex, detecting hardware vulnerabilities has become more challenge. Recently, hardware fuzzing techniques have shown promising results in generating complex programs for processor testing. However, the complexity of processors continues to limit the speed of vulnerability detection and the ability to achieve sufficient coverage.This paper introduces FeedbackFuzz, a novel processor fuzzer aimed at significantly accelerating vulnerability detection. FeedbackFuzz enhances detection efficiency by generating more effective test programs and optimizing the process of identifying vulnerabilities in longer programs. It designs a feedback engine for processors with out-of-order execution capabilities. Additionally, we leverage large language models (LLMs) to handle the complexity of locating vulnerabilities in lengthy programs, greatly speeding up the detection process. We evaluated FeedbackFuzz on several open-source RISC-V processors. Our evaluation demonstrates that the efficiency of FeedbackFuzz has increased by 40% compared to Cascade, with a coverage acceleration of 16.7%.