10019474MLighter: Shedding light on Machine Learning Blind-SpotsClosedCollaborative R&DInnovate UKDeveloping machine learning systems requires knowledge on multiple aspects such as data, features and algorithms that normally need professionals to guarantee a robust and sound solution. These systems are indeed powerful and potentially will massively benefit our societies, however, we can aim to further improve them. Improving machine learning demands a proper adaptation of the way we test software to be more specific for these systems. MLighter adopts a new strategy to test machine learning models; it follows a holistic approach based on security, functionality and performance testing and provides tools to identify blind-spots in machine learning. During the first weekend of its existence, MLighter exposed 78 new blind-spots in the form of software bugs in one of the major machine learning systems (10 of which are exploitable). These blind-spots are related to operational software failures that cost globally $1.56 trillion, according to studies. Unfortunately, some of the blind-spots might lead to life-threatening situations, such as when smart cars misread the traffic signs or delay their responses depending on the traffic conditions. These different problems, which go beyond the normal testing, were the main motive to build MLighter, as a novel solution for detecting blind-spots in machine learning systems at different levels. Our solution is very useful for coders and QA Testers, and it is currently detecting exploits affecting major ML systems. Indeed, the results of our prototype are very promising, therefore we are looking to join forces with investors and collaborators to build and scale up MLighter in order to perfect the implementation of ML technology globally.