A Reinforcement Learning Approach to Multi-Parametric Input Mutation for Fuzzing

Fuzzing is a software testing technique that involves providing a target program with unexpected or randomly generated input to observe its behavior. Traditional fuzzing techniques generate these inputs by randomly mutating predefined seed inputs. This random mutation technique operates without feedback from the target program, making it difficult to achieve high code coverage and detect vulnerabilities in the program. Reinforcement learning (RL) has emerged as a promising approach to address that limitation by learning the optimal mutation strategies. However, existing reinforcement learning-based fuzzing methods do not handle the mutation of multi-parametric test cases, where an input contains multiple parameters that require simultaneous yet distinct mutation strategies. This paper presents a reinforcement learning-based fuzzing technique that optimizes the mutation strategy for multi-parametric test cases. Unlike traditional approaches that apply a single mutation to generate the next input, our method mutates multiple parameters within a test case in a single mutation round, each with distinct mutation strategies. We evaluate the proposed method against traditional fuzzing techniques targeting a Power Generation Supervisory Control System, achieving a $70 \%$ code coverage and twice as many crashes. Additionally, we compare our approach with an existing reinforcement learning-based fuzzing method on two widely used targets (xmllint and tcpdump) and observe a $22 \%$ increase in code coverage for xmllint and a $43 \%$ increase for tcpdump, demonstrating substantial improvements in both code coverage and crash detection.