KCFuzz: Directed Fuzzing Based on Keypoint Coverage

Abstract

Directed fuzzing, as an efficient method to focus on a specific set of targets in the program, often works better than random fuzzing when combined with a researcher’s empirical judgment. However, the current directed fuzzing work is not efficient enough. In previous studies, some have generated closer seed inputs by guiding the execution path through the distance from the target region, but the distance guided algorithm is less robust. Some studies used selective symbolic execution for directed testing to alleviate the path explosion problem, but it brings a higher false-positive rate. In this paper, we propose a keypoint coverage-based fuzzing (KCFuzz) method, which extracts the keypoint list using a control flow graph, obtains the keypoint list coverage information through runtime instrumentation, calculates the test priority of the seeds based on the overall coverage and keypoint coverage using an energy scheduling algorithm, and continuously generates test inputs closer to the target according to the specified mutation strategy. On this basis, a hybrid testing framework is implemented, using keypoint coverage directed fuzzing to generate a seed queue covering keypoints, using offspring generation strategies and hybrid execution technology, and further exploring the new state of the program according to changes in overall and keypoint coverage. The experimental results show that the KCFuzz method can efficiently induce the generation of seed queues to reach the target region, and at the same time, the depth and validity of the exploration paths are higher than those of the most advanced directed fuzzing methods such as AFLGo.