Using LLMs to detect attack techniques & Understanding Deep Learning

The MLAI Meetup is a community for AI researchers and professionals which hosts monthly talks on exciting research. Our format is:

• 6:00 - 6:20: Socializing
• 6:20 - 6:40: Announcements and AI news
• 6:40 - 7:40: Talk(s) and Q&A
• 7:40 - 8:00 Networking
• 8:00: Head to the nearest pub for dinner

We have a great event this month, with a short talk from Javier and a long talk from Martin!

Martin Nguyen: "Using LLMs to detect attack techniques from threat intelligence reports"

Abstract: Cyber threat intelligence (CTI) reports have been one of the most important sources for analysts and researchers to prepare against potential threats and defend critical infrastructure, yet there have been multiple problems in automation of CTI analysis, due to text complexity.

In this talk, we will discuss the use of open-source Large Language Models to analyse this data. We also present a novel two-step pipeline to extract attack techniques from CTI reports, which was developed as collaboration between Swinburne University and CSIRO. This approach has significantly improved the extraction performance of LLMs, and with several attack techniques surpassing an F1-score of 0.90. Furthermore, it also shows potential for reliable automated CTI systems to support cybersecurity operations.

Speaker Bio: Martin is currently working as an AI Developer in [Chamomile.ai](http://chamomile.ai/), focusing on topic modelling. Prior to this work, he also did a Honours degree in Computer Science from Swinburne University, and had research experience with CSIRO's Data61 in applying large language models and natural language processing (NLP) methods for analysing cybersecurity data. His current interest lies on applying AI to solve real-world challenges, especially in cybersecurity.

Javier Candeira: "Books We Love: Simon Prince's 'Understanding Deep Learning' (2023)"

Blurb: Javier will present a review of an authoritative, accessible, and up-to-date technical book on what deep learning is all about: Simon Prince's Understanding Deep Learning.

Starting from first principles, this textbook builds an explanation of the mathematical workings of different of deep neural network architectures, starting from the single layer perceptron and arriving at cutting edge topics like transformers and diffusion models. Two chapters close the book with a change of pace and tone. "Ethics in deep learning" is presented as a literature review co-written with Philosopy Professor Travis LaCroix. The chapter "Why does deep learning work? is titled as a question; Prince argues that it's surprising that deep neural networks are easy to train and it's surprising that they generalise!

Like the book it's based on, the talk will touch on the "unreasonable effectiveness of deep learning", including discussion of memorisation and generalisation, double descent and 'grokking', lottery tickets and pruning, etc.

Speaker Bio: Javier Candeira is a software engineer, entrepreneur, public speaker, conference organiser, and a lifelong student of way too many topics, including machine learning.