Hayes, Jamie; (2021) Towards private and robust machine learning for information security. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text JamieHayes_PhD_Thesis.pdf Download (14MB) | Preview

Abstract

Many problems in information security are pattern recognition problems. For example, determining if a digital communication can be trusted amounts to certifying that the communication does not carry malicious or secret content, which can be distilled into the problem of recognising the difference between benign and malicious content. At a high level, machine learning is the study of how patterns are formed within data, and how learning these patterns generalises beyond the potentially limited data pool at a practitioner’s disposal, and so has become a powerful tool in information security. In this work, we study the benefits machine learning can bring to two problems in information security. Firstly, we show that machine learning can be used to detect which websites are visited by an internet user over an encrypted connection. By analysing timing and packet size information of encrypted network traffic, we train a machine learning model that predicts the target website given a stream of encrypted network traffic, even if browsing is performed over an anonymous communication network. Secondly, in addition to studying how machine learning can be used to design attacks, we study how it can be used to solve the problem of hiding information within a cover medium, such as an image or an audio recording, which is commonly referred to as steganography. How well an algorithm can hide information within a cover medium amounts to how well the algorithm models and exploits areas of redundancy. This can again be reduced to a pattern recognition problem, and so we apply machine learning to design a steganographic algorithm that efficiently hides a secret message with an image. Following this, we proceed with discussions surrounding why machine learning is not a panacea for information security, and can be an attack vector in and of itself. We show that machine learning can leak private and sensitive information about the data it used to learn, and how malicious actors can exploit vulnerabilities in these learning algorithms to compel them to exhibit adversarial behaviours. Finally, we examine the problem of the disconnect between image recognition systems learned by humans and by machine learning models. While human classification of an image is relatively robust to noise, machine learning models do not possess this property. We show how an attacker can cause targeted misclassifications against an entire data distribution by exploiting this property, and go onto introduce a mitigation that ameliorates this undesirable trait of machine learning.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item