Verification of semantic key point detection for aircraft pose estimation
Abstract
Accurate pose estimation systems are essential components for the development of fully autonomous aircraft. While neural networks form the backbone of accurate and versatile
implementations, their susceptibility to adversarial attacks may hinder the safe operation of the aircraft. To mitigate harmful behaviours, formal verification can help assess their adversarial robustness before deployment. In this talk I will outline the formal verification of an industrial-scale neural network built by Boeing and embedded within a pose estimation system. I will discuss the methods powering the underlying verification engine, identify the conditions under which the network can be verified, and summarise the verification results drawn with respect to white noise and photometric perturbations.
Speaker bio
Panagiotis Kouvaros is an Assistant Professor at the University of Limassol and a and Senior Research Engineer at Safe Intelligence. He works on the formal verification of Artificial Intelligence, covering both symbolic and connectionist systems. He is currently leading the
verification effort at Safe Intelligence, a spinout company of Imperial College London. He received his PhD inComputing, Imperial College London in 2016.
Practical
Speaker presents in person at Integreat, 8th floor, Niels Henrik Abels building.
Zoom: https://uio.zoom.us/j/68132593627
Organisers
- Ana Ozaki, anaoz@ifi.uio.no
- Maria Dikova, mariyad@uio.no