Abstract
Deep neural network(DNN) is the core of artificial intelligence applications. Recent progress in large-scale foundation models (FMs) further amplifies the importance of systematic reliability testing, as these complex models may fail unexpectedly under subtle input variations. Such challenges highlight the continued need for principled fuzzing techniques capable of discovering diverse execution behaviors and potential security problems in DNNs. However, despite recent progress in DNNs fuzzing, existing approaches remain limited in their execution-status analysis and mutation strategies, motivating our design of a more comprehensive and configurable framework. In this paper, we propose MCM-DeepFuzz, a fuzzing framework for deep neural networks with configurable execution status analysis and mutation strategies to improve the effectiveness of fuzzing. It combines 6 coverage criteria to analyze execution status and introduces 7 image operations as mutation strategies. Furthermore, we introduce the concept of sensitivity to evaluate these criteria with the MNIST dataset on CNN and MLP models, showing the characteristics and differences of existing coverage criteria. We also analyze how parameters and hidden layers affect the sensitivity of coverage criteria. Extensive experiment also demonstrates the feasibility and effectiveness of these mutation strategies.