pAFL: Adaptive Energy Allocation with Upper Confidence Bound

Recently, Fuzzing has regarded as one of the most widely used tools of discovering software vulnerabilities, due to its effectiveness and efficiency. With various fuzzers developing, ineffective seed generation has emerged as a concern. American Fuzzy Lop (AFL), a coverage-guided fuzzer, allocates mutation energy to seeds to create new inputs. Nevertheless, AFL's fixed mutation energy for the same seed after multiple mutations leads to the exploration of unproductive paths, reducing vulnerability detection efficiency. To overcome this problem, we proposed a novel adaptive energy allocation scheme, pAFL. Utilizing reinforcement learning, pAFL dynamically assigns energy to seeds in iterations. Initially, it assigns more energy to promising seeds which are judged by several native metrics, followed by employing the Upper Confidence Bound (UCB) algorithm to balance exploration and exploitation. This prevents the same seeds from over-exploitation and improves exploration among different seeds. The evaluations on LAVA-M dataset and 7 real-world programs demonstrate that pAFL outperforms AFL significantly. Additionally, we verifies that pAFL could achieve better performance by overcoming more path constraints on fuzzer_challenges dataset compared to AFL, AFLFast, EcoFuzz and MOPT.