Benedikt Brückner (2021 cohort) presented his paper, ‘A Semidefinite Relaxation based Branch-and-Bound Method for Tight Neural Network Verification’ at the 37 th AAAI Conference on Artificial Intelligence 2023 (AAAI 2023) in Washington DC.
Benedikt’s work focuses on neural network verification. As Benedikt explains, “Adversarial attacks have shown that neural networks are often fragile and small changes in the input of a network may cause unexpected and undesired changes of the output of the network, for example the classification it provides for a given input. This is why there has been a lot of work on neural network verification over the past years which aims to formally certify that a network is robust to those small input changes”.
Benedikt describes how his paper introduces a tight and efficient method for verifying the robustness of neural networks which is based on Semidefinite Programming techniques. Benedikt says that, “It improves the tightness of the approximations that are used to model the behaviour of the network in the verification process by adding a set of linear constraints based on an eigenvector decomposition to the problem. The approach is augmented with a Branch-and-Bound framework which allows us to gradually improve the approximations to the point where we are able to encode the network’s behaviour exactly. We are able to prove that this method yields tighter approximations than the state of the art, and also conduct an experimental evaluation in which it is able to outperform the baselines in terms of the verified accuracy”.
For Benedikt, the conference was an opportunity to expand his knowledge, gain new ideas and explore different areas of research through the variety of talks and presentations on offer. He says, “The tutorial sessions during the first days served as a good introduction to a number of topics while the paper presentations then delved deeper into recent progress in those areas. The poster sessions offered an opportunity to discuss topics in more detail and were also great for networking”. There was also time to take in the sights of Washington, and Benedikt enjoyed meeting fellow researchers from across the world.
Benedikt really enjoys being part of the STAI CDT and appreciates the cohort-based nature of the training. He says, “Having a number of like-minded people who work on similar things and might be facing similar problems definitely is a huge benefit since it is easy to discuss problems that one might encounter. Besides that, it is also great for getting to know people and socialising – I really enjoy the CDT events we have and there’s a lot going on within the cohorts and the CDT in general. There’s an amazing sense of community there and I’m definitely happy to have joined the CDT instead of doing a standard PhD and am very grateful for this opportunity!”.
Benedikt is second author on the paper, and his co-authors are Jianglin Lan and Professor Alessio Lomuscio.