Anomaly detection for security

Anomaly detection is the task of identifying items or events which deviate significantly from normal appearance or behaviour. This is a well-established approach in financial fraud detection, but is applicable in many other areas within the security domain. In X-ray screening of baggage and cargo it can be used to detect concealment, even if a threat is not directly recognisable. In biometric verification it can be used to detect tampering, such as facial morph images which match two identities. In home security it can enable advanced alarm systems that warn when unusual physical or digital activity is occurring.

Anomaly Detection is a sub-problem of machine learning. It is common for machine learning applications to be hindered by a lack of sufficient available training data. This problem is raised to its most extreme form in anomaly detection, where there may be no available example anomalies; indeed this is entirely the point, anomaly detection is (roughly) about detecting known unknowns, rather than known knowns.

The answer, to be explored in this PhD project, is to use methods of self-supervised rather than supervised learning. Self-supervised learning using proxy tasks which can be defined on normal data, allowing effective data representations to be learnt from that data alone, rather than on the contrast between normal and threat data. Effective representations are they to effective modelling of the distribution of normal data so that anomalous deviations can be spotted.

The general goal of the PhD will be to develop self-supervised learning methods that are effective for anomaly detection. In particular to develop an understanding of how the proxy task should relate to the type of anomaly to be detected. Specific goals will be to pioneer anomaly detection methods in different security/crime domains, dependent on access to sufficient (large) quantities of normal training data, and suitable smaller datasets of anomalies that can be used to test the resulting methods.

The first problem to be worked on will be anomaly detection in X-ray security imaging. Here ample data is available for training and testing, and a particular feature of the data (dual-view images) provides the opportunity for an effective proxy task - to wit, learning representations that allow the two views of a bag or parcel to be paired. Other problems that we plan to explore are anomaly detection in audio (for deep fake detection), in video streams (as an AI-driven flexible home emergency/crime alarm), and in text (for detecting phishing, and similar, emails).