Congratulations to Patrick Henriksen of the 2019 cohort of the UKRI Centre for Doctoral Training in Safe and Trusted AI, whose paper has been accepted for publication and presentation at the 24th European Conference on Artificial Intelligence (ECAI 2020). Originally scheduled for June, Patrick is now due to present the paper, co-authored with his supervisor (and CDT Deputy Director) Professor Alessio Lomuscio, at the rescheduled conference in August, due to COVID-19.
In the paper, entitled Efficient Neural Network Verification via Adaptive Refinement and Adversarial Search, a novel formal verification algorithm for neural networks is proposed. Neural networks have had an enormous impact in the field of machine learning over the past few years; however, recent work has shown that they are inherently fragile. For example, in image classification tasks, adding a tiny amount of noise to the input image can lead to misclassification, in turn may have catastrophic consequences in safety-critical applications such as autonomous driving and avionics. It therefore remains of utmost importance to develop methods that can formally verify the correctness of neural networks. In this paper, the authors propose a novel verification algorithm for local robustness properties in nerual networks. More specifically, their approach consider classification networks and the algorithm aims to prove that the netowrk’s classification does not change for a set perturbation of the original image. They achieve state-of-the-art performance by extendng Symbolic Interval Propogation-based techniques with several novel improvements, such as a local search counter-examples and improved branch and bound phase.
ECAI 2020 is Europe’s leading research conference for artificial intelligence. This year, the conference takes place in Santiago de Compostela in Spain, and is supported by organisations such as Huawei, Microsoft and Accenture.