George Kampanos (MSc Cyber Security student)
Our analysis shows that although more than 60 per cent of websites store third-party cookies in both countries, only less than 50 per cent show a cookie notice and hence a substantial proportion do not comply with the law even at the very basic level. We find only a small proportion of the surveyed websites providing a direct opt-out option, with an overwhelming majority either nudging users towards privacy-intrusive choices or making cookie rejection much harder than consent.
Our results differ significantly in some cases from previous smaller-scale studies and hence underline the importance of large-scale studies for a better understanding of the big picture in cookie practices.
Determining the Acceptability of Abstract Arguments with Graph Convolutional Networks
Abstract Argumentation is a framework for non-monotonic reasoning that models situations involving conflict using a minimal representation. This "calculus of conflicts" has many applications both related to argumentation in language and more broadly to scenarios where you need to make sense of conflicting information such as in sensor networks or intelligence analysis. Unfortunately, most problems in this field fall into intractable computational classes and it is easy to devise problems that due to their size and complexity will be unsolvable using exact algorithms. My research, therefore, seeks to create an approximate solution approach using Graph Neural Networks that will have sufficiently high accuracy and speed to enable them to be used in Cloud-scale applications. Thereby, it will become possible to apply this mode of reasoning to a much larger set of real-world problems than is currently feasible.
An example abstract argumentation system where a-f are arguments, arrows represent attacking relations, e.g. a-->b means argument a attacks argument b.
Learning good encodings for translating constraint satisfaction problems into boolean satisfiability
Felix Ulrich-Oltean (PhD student)
An explosion on board Apollo 13 led to critical loss of oxygen and fuel. The crew and ground control had to design a new sequence of actions to help the astronauts return to Earth safely without blowing the electrical circuits or asphyxiating. This extreme planning is just one example of a constraint satisfaction problem (CSP).
Harsh constraints and competing priorities have become a fact of life for many during the last few months. Constraint programming is already used in applications like timetabling, logistics and verifying hardware circuits. This approach has much to offer as we navigate new realities, such as operating complex systems with limited energy resources, scheduling medical teams in COVID-secure ways or placing 5G masts to reach all intended users.
Several languages exist for expressing a constraint problem in an exact but human-friendly way. Once a problem is thus "modelled", constraint solvers can find suitable solutions without being given any instructions.
One popular approach to solving a constraint satisfaction problem is to translate it into a logical statement using only True/False variables, along with the operators NOT, AND and OR. The problem is then solved by a Boolean Satisfiability (SAT) solver - a type of software that has seen impressive advances in recent years.
CSPs can be translated to SAT in different ways. My work uses machine learning techniques to improve this process by predicting good choices when encoding different aspects of the problem. Better encoding choices lead to faster solving times, making more problems tractable.