基于序列生成对抗网络的智能模糊测试方法

Intelligent Fuzzy Testing Method Based on Sequence Generative Adversarial Networks

漏洞的数量增长以及超危、高危等高度危害漏洞的大量出现使得网络安全形势面临着极大挑战,模糊测试作为主流的安全测试手段被广泛应用.测试用例生成作为核心步骤直接决定了模糊测试效果的优劣,然而传统的基于预先生成、随机生成以及变异策略的测试用例生成方法面临着覆盖面低、人工成本高、质量低下等瓶颈问题,如何生成高质量、高可用、完备的测试用例是智能模糊测试的难点问题.针对于此问题,提出一种基于序列生成对抗网络(SeqGAN)模型的智能模糊测试方法,结合强化学习的思想将测试用例生成抽象为普适的非定长离散型序列数据的学习和近似生成问题,创新性地在生成器部分增加可配置的嵌入层来规范生成,并采用动态权重调整的方式从真实性和多样性2个维度设计奖励函数,最终实现自动化、智能化地构造全面、完备、可用的测试用例集,以达到灵活、高效的智能模糊测试的目标.从有效性和通用性2个层面分别对所提方案进行了验证,在4种不同测试目标下平均95%以上的测试用例通过率以及平均10%的目标缺陷检测能力充分证明了方案的通用性,在4种不同方案对比下98%的测试用例通过率、9%的目标缺陷检测能力以及单位时间内2万条可用测试用例的生成能力充分证明了方案的有效性.

The increase in the number of vulnerabilities and the emergence of a large number of highly dangerous vulnerabilities,such as supercritical and high-risk ones,pose great challenges to the state of network security.As a mainstream security testing method,fuzz testing is widely used.Test case generation,as a core step,directly determines the quality of fuzz testing.However,traditional test case generation methods based on pre-generation,random generation,and mutation strategies face bottlenecks such as low coverage,high labor costs,and low quality.Generating high-quality,highly available,and comprehensive test cases is a difficult problem in intelligent fuzz testing.To address this issue,this paper proposes an intelligent fuzz testing method based on the sequence generation adversarial network(SeqGAN)model.By combining the idea of reinforcement learning,the test case generation is abstracted as a learning and approximate generation problem for universally applicable variable-length discrete sequence data.Innovatively,a configurable embedding layer is added to the generator part to standardize the generation,and a reward function is designed from the dimensions of authenticity and diversity through dynamic weight adjustment.This ultimately achieves the goal of automatically and intelligently constructing a comprehensive,complete,and usable test case set for flexible and efficient intelligent fuzz testing.This paper verifies the proposed scheme from two aspects of effectiveness and universality.The average test case pass rate of over 95%and the average target defect detection rate of 10%under four different testing targets fully demonstrate the universality of the scheme.The 98%test case pass rate,9%target defect detection rate,and the ability to generate 20000 usable test cases per unit time under four different schemes fully demonstrate the effectiveness of the scheme.