Assessing the reliability of machine learning cloud services through fault injection

Abstract

Building Machine Learning (ML) systems can be tricky, and ML technologies are known to have a steep learning curve. Those difficulties led to the popularization of ML services. Although many recent studies have addressed the vulnerability of ML models to target attacks, not enough attention has been directed to the effect of typical data faults on the reliability of ML services. Such faults may originate in the applications that rely on ML services, caused by hardware or connection failures, bugs, or undefined behaviour. Consequently, those faults can be reflected in the data produced by such applications and sent to the ML services. Seeking to evaluate the reliability of these services and focusing on the Computer Vision (CV) domain, this work presents an empirical study on the injection of common data faults into the data sent to CV services. The results from 11 commercial CV services indicate that data faults may significantly affect the reliability of the CV services, as evidenced by misclassification rates ranging from 14% to up to 63%. On the other hand, data faults do not seem to have the same impact on the fairness of CV services, even though some fault configurations lead to significant fairness impacts.