AIAI Seminar - 27 February 2023 - Talks by Nick Ferguson, Jonathan Feldstein and Tarini Saka

Speaker:   Nick Ferguson

Title:          An Outline and Discussion of Selected Issues in the EU AI Act

Abstract: 

The European Union's proposed regulation on artificial intelligence (the AI Act) outlines the legal responsibilities of AI system developers in terms of transparency, risk management, and explainability. It is an ambitious proposal, yet not without flaws. Given the cross-disciplinary collaboration required to produce such a document, it is inevitable that some conceptual misunderstanding will occur. Sometimes, concepts appear to be defined differently to the way they are defined in AI literature, while other times, there is simply a lack of a clear definition. Additionally, the Act also requires AI providers to prove conformity to the regulation, demonstrating necessary levels of transparency. In this talk, I will address these where conceptual understandings do not agree, technological challenges raised by such disagreement, difficulties with demonstrating conformity to the Act.

Speaker:   Jonathan Feldstein

Title:           Principled and Efficient Motif Finding for Structure Learning in Lifted Graphical Models

Abstract:

Structure learning is a core problem in AI central to the fields of \textit{neuro-symbolic AI} and statistical relational learning. It consists in automatically learning a logical theory from data. The basis for structure learning is mining repeating patterns in the data, known as structural motifs. Finding these patterns reduces the exponential search space and therefore guides the learning of formulas. Despite the importance of motif learning, it is still not well understood. We present the first principled approach for mining structural motifs in lifted graphical models, languages that blend first-order logic with probabilistic models, which uses a stochastic process to measure the similarity of entities in the data.

Our first contribution is an algorithm, which depends on two intuitive hyperparameters: one controlling the uncertainty in the entity similarity measure, and one controlling the softness of the resulting rules. Our second contribution is a preprocessing step where we perform hierarchical clustering on the data to reduce the search space to the most relevant data. Our third contribution is to introduce an O(n ln n) (in the size of the entities in the data) algorithm for clustering structurally-related data. We evaluate our approach using standard benchmarks and show that we outperform state-of-the-art structure learning approaches by up to 6\% in terms of accuracy and up to 80\% in terms of runtime.

Speaker:   Tarini Saka    

Title:          A Collaborative Human-AI Approach to Tackle Phishing Attacks

Abstract: 

Phishing is one of the most prevalent and dangerous types of cybercrime currently faced by organizations and individuals worldwide. Previous research has emphasized the importance of identifying email campaigns to combat phishing attacks more quickly and effectively. However, there is a lack of understanding of what a phishing campaign is and the common attributes that campaign emails share. To address this, we present a phishing codebook, designed through iterative coding specifically to extract information from phishing emails. Our goals are two-fold: to create a dataset to understand the qualitative characteristics of phishing campaigns, and to find a systematic way to extract important information from a phishing email to create a more structured representation. To the best of our knowledge, this is the first attempt to create such a phishing codebook. It aims to standardize email text and improve clarity and conciseness. I will present our coding approach, goals and future directions.