About the project
Machine learning (ML) components such as neural networks (NNs) have become pervasive in complex system design. However, NNs are known to be fragile due to adversarial attacks.
Only very recently, solvers were developed, such as Marabou or alpha-beta-Crown, that can verify NNs of several thousand nodes. This means that they stretch to medium-sized NNs used in Interconnected Cyberphysical Systems (ICS).
Until recently, 3 problems prevented their practical adoption:
- lack of convenient user interfaces in which complex property specifications for NNs could be written
- poor integration of these solvers into richer provers where analysis or verification of the overall complex system behaviour takes place
- lack of efficient ways to automatically repair NNs when their verification fails
The research team led by Katya Komendantskaya has implemented a tool, Vehicle, that uses types and differentiable logic to solve these problems.
The aim of this PhD project is to revolutionise the ways ICS embed and secure NN components. The project will explore the range of NN-related vulnerabilities in ICS, both in terms of fragility/vulnerability to attack and repair, cast as “property specification”, and “property-driven training” in Vehicle's terms.
You will have a chance to engage with researchers at the National Robotarium in Edinburgh, and industrial collaborators such as Imandra or Schlumberger, and will be encouraged to seek new industrial links.
This project is funded through the UKRI MINDS Centre for Doctoral Training. This is one of 16 PhD training centres in the UK with a unique focus on advancing AI techniques in the context of real-world engineered systems. Its remit spans novel hardware for AI, AI and machine learning, pervasive systems and IoT, and human-AI collaboration. We provide enhanced cross-disciplinary training in electronics and AI, entrepreneurship, responsible research and innovation, communication strategies, outreach and impact development as part of an integrated 4-year iPhD programme.
