Collaborators Sanmi Koyejo, UIUC
Real-time systems are often used to control physical systems and hence, designed to provide safety guarantees — often by ensuring that tasks complete before their deadlines. This makes critical (often safety-critical, high priority) tasks valuable targets for an adversary, as a disruption to these tasks could result in serious damage to the system, operators and even the environment. Hence, knowledge of the timing properties of such tasks (especially their execution schedules) is important knowledge that can be used by adversaries to launch targeted attacks. Scheduler-based Side-channel Attack (ScheduLeak) assumes that the adversary has the information (number of tasks, period, execution) about the task set a priori. The main goal of the project is to relax the assumption by creating a predictor model that can infer the number of tasks, execution of past tasks, and the prediction of the future execution of the tasks using deep learning. We also investigate the behaviour of the predictor model and see how the prediction problem can further our understanding of the recurrent neural nets.