اكرم كلائي

Mathematical modeling of physical system behavior forms the foundation for constructing most cyber-physical systems (CPS). These mathematical models, composed of a set of coherent equations, describe specific behaviors and properties of the system. However, some influential variables may remain hidden due to a lack of awareness or understanding by system designers, which can lead to modeling errors and inaccurate predictions. Leveraging machine learning models to develop predictive models that perform reliably under novel conditions presents a promising alternative. Yet, the black-box nature of such models makes it challenging to predict and validate their behavior in unforeseen scenarios. These challenges render machine-learning-enabled CPS—especially in safety-critical domains—susceptible to weaknesses in testability and explainability. A review of the literature reveals a predominant focus on black-box testing approaches, mainly due to the weak correlation between white-box coverage criteria and fault detection effectiveness. These approaches often search for diverse test data that violate functional or safety requirements but tend to overlook the internal decision-making logic of the system in response to environmental events. In reality, the problem of testing and explaining behavior in such complex systems is a domain analysis problem and must address domain-specific root causes of failure. Domain analysis partitions the system's input space into distinct subdomains based on software specifications, assigning each subdomain to a particular behavior or execution path. This helps identify and characterize the system’s behavior under varying conditions. The primary goal of this dissertation is to leverage domain analysis in tandem with search-based testing techniques to enhance the detection and explanation of incorrect behaviors in learning-enabled CPS. A multilayered and coherent framework is proposed to address testability and explainability across various levels of the system—from deep learning models to high-level decision logic. The proposed framework comprises four complementary steps. The first step focuses on deep neural networks (DNNs) with predictive tasks. By integrating domain analysis and decision logic extraction, a novel test data generation criterion is introduced. Framed as a multi-objective optimization problem, this method enables the identification of edge and critical test cases that challenge model performance in previously unseen scenarios. The second step targets classification-based DNNs and proposes a new metric for estimating the collective influence of neurons on misclassifications. By combining spectrum-based fault localization and multi-stage gradient ascent, this method generates new datasets that improve the accuracy and clarity of fault root-cause explanations. The third step addresses rule-based decision-making components by introducing a behavior-oriented criterion to identify critical decision paths. As in the previous steps, this is formulated as a multi-objective optimization problem, enhancing the identification of behavioral conditions that may cause system failures in sensitive decisions. The fourth step focuses on DNN-based decision-making modules. It introduces a novel risk metric to identify critical regions of the input space. The resulting optimization process generates diverse test cases that enable a more precise characterization of the system’s behavioral space. Test execution time is also reduced using a surrogate model based on ensemble reinforcement learning. To eva‎luate the effectiveness of the proposed approaches and compare them with baseline methods, extensive experiments were conducted on a set of DNN models and the source code of multiple cyber-physical systems in the autonomous driving domain. eva‎luation metrics included diverse indicators of testability and explainability, using publicly available and standardized datasets. The results demonstrate significant improvements in testability (up to 65%) and explainability (up to 69%) over baseline methods. Statistical analyses further confirm the significance of the observed differences.

سيستم‌هاي سايبر- فيزيكي , يادگيري ماشين , توضيح‌پذيري , آزمون نرم‌افزار , مكان‌يابي خطا , تحليل دامنه , توليد داده آزمون

Cyber-physical systems , Machine learning , Explainability , Software testing , Fault localization , Domain analysis , Test data generation

Akram Kalaee

Saeed Parsa