基于形式化可解释人工智能的网络入侵检测方法  

Towards Rigorous Explanations for Neural Network Driven Intrusion Detection

作　　者：周倩如 ZHOU Qianru(School of Computer Science and Engineering,Nanjing University of Science and Engineering,Nanjing 213000,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京213000

出　　处：《智能安全》2024年第3期34-44,共11页Artificial Intelligence Security

基　　金：国家科技部项目(G2022182006L)。

摘　　要：随着人工智能的火速发展,神经网络被广泛应用到各行各业,但是神经网络模型的不可解释限制了其所做出决策的可信度。目前的可解释机器学习方法大都只能提供模糊的、近似的解释,在涉及安全相关的重要领域,每一个字节的解释偏差都可能引发重大误导,可是逻辑严谨、精确的可解释方法会涉及组合爆炸,非常复杂。提出了计算边界值(BIV)算法,一种计算模型决策恒真域和恒假域之间边界的算法。首先,基于模型的权重、偏置和激活函数性质,通过知识编译将模型表示成形式化逻辑表达式——基于阈值的线性函数。然后,通过将模型的决策过程看成是二元对抗博弈,提出恒真域和恒假域的概念,并在此基础上,将逻辑严谨的可解释问题转化成计算恒真域与恒假域边界的问题。随后,提出了直接计算恒真域与恒假域之间边界的算法——BIV算法,可以直接精准计算出形式化解释,即模型的决策规则。最后,在真实DoS网络入侵检测场景上,对BIV算法进行测试,并将结果与现有可解释方法SHAP进行了分析和对比。结果表明,BIV算法不仅可以提供精准严谨的解释,而且在执行效率上有明显优势。With the pervasion of the artificial intelligence,neural network models have been adopted widely for decision making.However,the decision they made cannot be trusted because they cannot explain the reasons.Most of the current explainable AI methods can only provide vague and ambiguous explanations,which cannot satisfy the requirement in security-concerned domains,such as cybersecurity,where the misinterpretation of even one bit can cause tremendous semantic misunderstandings.This paper proposes a rigorous explanation methodology,boundary input value(BIV)algorithm,based on knowledge compilation.Neural network models are presented into formal logic expressions,then prime implicants are extracted from them,which is the rigorous explanations for the model.Proposed BIV algorithm has been verified on a real-life DoS intrusion detection neural network model.Explanations extracted and time overhead are analyzed and compared with current explanation method SHAP.It is evident that the proposed BIV algorithm can provide rigorous explanations efficiently,and thus is scalable.

关 键 词：可解释人工智能 知识编译 神经网络 入侵检测 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]