AdapFuzzer: An Adaptive Fuzzing Framework for DNNS with Diversity-Aware Seed Selection and Gan-Based Mutation

Abstract

Deep learning systems, built on data-driven learning paradigms, are highly sensitive to input perturbations, where even slight variations can lead to substantial output deviations or incorrect decisions, posing serious reliability and safety concerns. To address these challenges, we propose AdapFuzzer, an adaptive fuzzing framework that incorporates feedback-driven control principles into the testing process. AdapFuzzer consists of three core components: (1) NFuzzer, a diversity-driven seed selection module that iteratively constructs a representative seed pool using deep feature-based dissimilarity measurement; (2) FAGAN, a mutation engine that leverages generative adversarial learning to produce high-quality and diverse adversarial variants; and (3) a test management module that continuously analyzes coverage and fault-triggering feedback to refine both seed prioritization and mutation strategies. Through this closed-loop optimization, AdapFuzzer dynamically adapts its testing behavior to the evolving detection capability of seeds and mutation operators. Experimental results on multiple deep learning models demonstrate that AdapFuzzer significantly improves testing effectiveness, achieving higher coverage and uncovering more erroneous behaviors than state-of-theart fuzzing approaches, thereby enhancing the robustness and security of deep learning systems.