DELFASE: A Deep Learning Method for Fault Space Exploration

Abstract

Cyber-Physical Systems (CPSs) are increasingly used in various safety-critical domains; assuring the safety of these systems is of paramount importance. Fault Injection is known as an effective testing method for analyzing the safety of CPSs. However, the total number of faults to be injected in a CPS to explore the entire fault space is normally large and the limited budget for testing forces testers to limit the number of faults injected by e.g., random sampling of the space. In this paper, we propose DELFASE as an automated solution for fault space exploration that relies on Generative Adversarial Networks (GANs) for optimizing the identification of critical faults, and can run in two modes: active and passive. In the active mode, an active learning technique called ranked batch-mode sampling is used to select faults for training the GAN model with, while in the passive mode those faults are selected randomly. The results of our experiments on an adaptive cruise control system show that compared to random sampling, DELFASE is significantly more effective in revealing system weaknesses. In fact, we observed that compared to random sampling that resulted in a fault coverage of around 10%, when using the active and passive modes, the fault coverage of DELFASE could be as high as 89% and 81%, respectively.

Towards Formal Fault Injection for Safety Assessment of Automated Systems

“…The challenge lies in identifying the optimal set of test instances that effectively analyze a system's dependability. Numerous testing methods have been proposed to address this challenge, including probabilistic approaches [20], coverage-based techniques [10], and heuristic as well as machine learning-based methods [29,28,35,36]. However, most of these methods rely on some level of prior knowledge about the target system.…”

Automating Fault Test Cases Generation and Execution for Automotive Safety Validation via NLP and HIL Simulation

“…In [70], Generative Adversarial Networks (GANs) and active learning are used to provide a fault injection testing technique. Causing safety failures entails the automatic generation and selection of faults.…”