SCADNet - Structural Causal Analysis of Deep Neural Networks

SCADNet - 深度神经网络的结构因果分析

• 批准号: 523877739
• 负责人: Professor Dr. Stefan Leue
• 金额: --
• 依托单位: Arbeitsgruppe Software und Systems Engineering
• 依托单位国家: 德国
• 项目类别: Research Grants
• 资助国家: 德国
• 项目状态: 未结题
• 来源: https://gepris.dfg.de/gepris/projekt/523877739?language=en
• 关键词: SCADNet; Structural; Causal; Analysis; Deep

Causal explanations based on counterexample computation and counterfactual causal reasoning are a cornerstone of the design and engineering of (Safety-)Critical Software-Driven Systems (CSS). They are the foundation of system debugging during iterative system design cycles, form the basis of failure forensics in litigation and are central to the certification of system safety as well as societal acceptance. In this project we focus on CSS that rely on decision-making components which are implemented using Machine Learning (ML) technology, in particular DeepLearning (DL). As widely acknowledged, obtaining explanations for this type of CSS is fraught with significant challenges, largely due to the absence of easily interpretable logical code structure and, as a consequence, the non-applicability of existing software verification technology. We will develop novel structural abstraction and analysis techniques for Deep Neural Networks (DNNs) that allow for the computation of counterexamples in the context of specification-based safety verification. We will then address the elicitation of counterfactual evidence for safety violations and next turn to proposing automatic repairs for DNNs that violate specified safety properties, thus contributing to the explanation of the behavior of a DNN. Finally, we will place these results in the context of a general theory of counterfactual causal analysis for ML-based CSS.

基于反例计算和反事实因果推理的因果解释是（安全）关键软件驱动系统（CSS）设计和工程的基石。它们是迭代系统设计周期中系统调试的基础，是诉讼中故障取证的基础，也是系统安全认证和社会认可的核心。在这个项目中，我们专注于依赖于决策组件的CSS，这些组件是使用机器学习（ML）技术实现的，特别是深度学习（DL）。如广泛承认的，获得对这种类型的CSS的解释充满了重大挑战，主要是由于缺乏易于解释的逻辑代码结构，因此，现有软件验证技术的不适用性。我们将为深度神经网络（DNN）开发新的结构抽象和分析技术，允许在基于规范的安全验证的背景下计算反例。然后，我们将讨论安全违规的反事实证据的引出，接下来我们将提出违反指定安全属性的DNN的自动修复，从而有助于解释DNN的行为。最后，我们将把这些结果的背景下，反事实的因果分析ML为基础的CSS的一般理论。