Research topic
The recent breakthroughs in Machine Learning applications, especially in Deep Neural Networks (DNNs), have caused significant progress in image classification and speech recognition applications. Autonomous driving is arguably one of the most important final applications making use of DNNs. Thanks to the use of these networks a vehicle can interpret what is happening around, the traffic signs and even the objects and people that are in the range of vision of the vehicle.
Nevertheless, Computer Vision is one of the toughest problems in Artificial Intelligence. Perceiving the surroundings accurately, quickly and energy efficiently is one of the most essential and challenging tasks for autonomous systems such as self-driving cars. Besides this real-time recognition requirement, recent works have proved that the output of DNNs can easily be fooled by adding relatively small perturbations to the input vector. And, what is more, with the irruption of Generative Adversarial Networks (GAN) this process of mimicking reality can even be automated. For instance, in a different scenario, Christie’s sold a portrait for $432,000 that had been generated by a GAN, based on open-source code written by Robbie Barrat of Stanford. Although not related with self-driving, this case gives an idea about the importance of reinforcing security in DNNs.
The proposal would then be focused on devising a real-time platform able to deal with the aforementioned issues and, regarding DNN security, being able to detect and counteract these threats.