Abstract

The growing adoption of DNNs in cloud systems necessitates performance analysis tools to identify potential issues with these learning-enabled systems prior to deployment. Existing analyzers either require modeling all the components faithfully (a white-box approach) or use inefficient black-box search. In this paper, we describe a gray-box alternative that does not require components to be modeled exactly yet is faster and finds substantially worse adversarial inputs than prior work. Using this method, we show that a state-of-the-art learning-enabled traffic engineering pipeline can underperform the optimal by $6×$ — a much higher number than what its authors were able to find. We conclude with a discussion of the challenges in generalizing this approach to other learning-enabled systems.

BibTeX Citation

Coming soon!