摘要
基于覆盖引导的模糊测试(Coverage-Guided Fuzzing,CGF)广泛应用于深度神经网络(DNN)测试,通过生成覆盖率高的衍生测试样本以提升模型的测试充分性.已有CGF方法侧重覆盖率和对抗成功率,较少考虑样本的多样性和模型鲁棒性提升.本文提出了一种基于多样性引导的深度神经网络测试用例生成方法DeepGA,旨在有限时间内生成多样化的测试用例,并评估和提升模型的鲁棒性.该方法从训练集中提取代表性的图像特征分布构建聚簇中心,以最大化多样性轮廓系数作为目标函数,通过遗传算法和模糊测试的思想对初始种子集群采用真实的图像变异策略,迭代生成与不同类别在特征分布上相似的测试用例图像.为了验证所提方法的有效性,本文基于4个DNN模型和3种不同的数据集进行了实验,结果表明DeepGA可以生成多样化的测试用例,并且生成测试用例可用于重训练以进一步提高被测模型的鲁棒性.在鲁棒性提升方面,与基于6种覆盖引导的测试用例生成方法对比,准确率最高提升了11.44%.
The coverage-guided fuzzing(CGF)is widely used in testing deep neural networks(DNN),generating derived test samples with high coverage to improve model testing sufficiency.Existing CGF methods mainly focus on coverage and adversarial success rate,with less attention to sample diversity and model robustness.This paper proposes DeepGA,a diversity-guided test case generation method for DNN,aiming to generate diverse test cases in limited time and enhance model robustness.The method extracts representative image feature distributions from the training set to construct cluster centers,maximizing the diversity silhouette coefficient as the objective function.It uses a genetic algorithm and fuzzing strategy with real image mutation tactics to iteratively generate test case images similar to different categories in feature distribution.Experiments on four DNN models and three datasets demonstrate that DeepGA generates diverse test cases that can be used for retraining to improve model robustness.Compared to six coverage-guided test case generation methods,the accuracy improves by up to 11.44%.
作者
苏祥
杨志斌
周勇
张海
SU Xiang;YANG Zhibin;ZHOU Yong;ZHANG Hai(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Key Laboratory of Safety-critical Software,Ministry of Industry and Information Technology,Nanjing 211106,China;Shanghai Institute of Satellite Engineering,Shanghai 201109,China)
出处
《小型微型计算机系统》
北大核心
2026年第1期181-192,共12页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62072233,U2241216)资助
航空科学基金项目(201919052002)资助。
