ارائه ­دهنده:

اکرم کلائی

  استاد راهنما:

دکتر سعید پارسا

  هیات داوران:

دکتر علی معینی

دکتر امید بوشهریان

دکتر بهروز مینایی بیدگلی

دکتر حسن نادری


 زمان 18 اسفند  ماه 1403

  ساعت: : 11:00 صبح
 

مکان: دانشکده کامپیوتر، طبقه سوم، اتاق سمینار

 

 چکیده پایان نامه :
 

Abstract

Mathematical modeling of the behavior of physical systems forms the foundation for constructing most cyber-physical systems. A mathematical model, through a set of consistent equations, describes the specific behaviors and characteristics of the system. However, some influential variables may remain hidden due to the designers' lack of awareness or insufficient understanding, leading to errors in the model and incorrect predictions. Using machine learning models to develop models that provide reliable predictions under new conditions is one solution. However, due to the black-box nature of these models, predicting and verifying their behavior under unexpected conditions is challenging. These challenges have rendered machine learning-based cyber-physical systems particularly problematic in terms of testability and explainability, especially in safety-critical applications. A review of the literature shows that research has focused on black-box methods due to the weak relationship between white-box coverage criteria and fault detection. These methods seek diverse test data and violations of functional/safety requirements, while ignoring the system's decision-making logic in response to environmental events. The issue of testing and explaining behavior in these complex systems is a domain-driven problem, and it is crucial to identify the root causes of domain errors. In domain analysis, the system's input space is partitioned into distinct domains based on software specifications, with each domain allocated to a specific path or behavior. This analysis helps identify and define the system's features and behaviors under various conditions. The main objective of this dissertation is to employ domain analysis in conjunction with search-based testing techniques to improve the discovery and explanation of incorrect behaviors in machine learning-based cyber-physical systems. The proposed approaches focus on identifying critical test paths based on the system's level of interaction with the environment and its involvement in decision-making logic, considering the riskiness of different input areas during test data generation, and covering the behavioral range of critical paths. Each proposed method is designed for a specific test artifact within machine learning-based cyber-physical systems. The first method, based on domain analysis and the model's decision-making logic, proposes a new criterion for generating test cases. This method focuses not only on data diversity but also on how the model reacts to inputs and interacts with the environment. This approach enables the discovery of edge and critical test cases that can challenge the model's performance under unforeseen conditions. The second method focuses on testing deep neural networks for classification tasks and introduces a new criterion for assessing the collective effect of neurons on the model's unsuccessful output. Using spectrum-based fault localization techniques and a multi-step gradient ascent method, this approach generates a new dataset that improves the accuracy of identifying and explaining the model's root causes of failure. The third method tests rule-based decision-making components by designing a behavior-based criterion to identify critical paths. By converting the issue of generating test cases into a multi-objective optimization problem, this method enhances the efficiency of determining the system's behavioral conditions and helps identify potential flaws in critical decision-making scenarios. The fourth method focuses on testing decision-making components based on deep neural networks and introduces a new criterion, the risk metric, for generating test data that identifies high-risk input areas. By turning the generation of test data into a multi-objective optimization problem, this approach produces a diverse set of test cases that aid in better describing the system's behavioral conditions. The time overhead for running the tests is also reduced using a collective-reinforcement surrogate model. To evaluate the performance of the proposed approaches and compare them with baseline methods, extensive experiments were conducted on a set of deep neural network models and the source code of several cyber-physical systems in the domain of autonomous vehicles. The evaluation criteria included various metrics for testability and explainability, using validated and standard datasets. The results show that the proposed approaches have significantly improved testability (by up to 65%) and explainability (by up to 69%) compared to baseline methods. Statistical analyses also confirm the significant difference between the proposed and baseline methods.