Virtualized-Fault Injection Testing: A Machine Learning Approach

We introduce a new methodology for virtualized fault injection testing of safety critical embedded systems. This approach fully automates the key steps of test case generation, fault injection and verdict construction. We use machine learning to reverse engineer models of the system under test. We use model checking to generate test verdicts with respect to safety requirements formalised in temporal logic. We exemplify our approach by implementing a tool chain based on integrating the QEMU hardware emulator, the GNU debugger GDB and the LBTest requirements testing tool. This tool chain is then evaluated on two industrial safety critical applications from the automotive sector.

I. Introduction

Fault simulation and fault injection (FI) are widely accepted techniques for assessing software robustness and error handling mechanisms of embedded systems [1]. In the automotive industry, FI has become a common practice to improve embedded systems quality and avoid the costs associated with untested safety-critical software [2]. The automotive safety standard ISO 26262 [3] recommends FI testing for electronic control unit (ECU) software at automotive software integrity levels (ASILs) C and D. Automotive recall cases such as Honda [4], Toyota [5] and Jaguar [6] show that in some extreme environmental conditions ECU hardware malfunctions can potentially give rise to ECU software faults. This highlights the importance of FI to avoid the costs and risks associated with unprotected software.