• Sorted by Date • Classified by Publication Type • Classified by Topic • Grouped by Student (current) • Grouped by Former Students •
Peta Masters, Daniel Gallagher, Gal A. Kaminka, and Mor Vered. Generation of Suspicious Behavior. In Proceedings of the ICAPS Workshop on Planning and Reasoning About Beliefs, Goals, and Intentions (PR-BGI), 2026.
(unavailable)
Patterns of behaviour are regarded as suspicious when they deviate from what is normal or expected for a given domain, where the motive is unknown and suggestive of risk. Recognising suspicious behaviour is valuable because it enables effective threat detection and response but in order to train machine learning models and security personnel to identify such behaviour, we must first define it and provide examples from which they can learn. Suspicion is a human construct, a property of the observer rather than of the behaviour itself, therefore we base our approach on goal recognition, which focuses intrinsically on the observer's perspective. We define and demonstrate several types of suspicious behaviour, synthesising plans for loitering, obfuscation, and deception. These reflect the behaviour of an agent that believes it may be being observed, and acts to conceal its intent. We evaluated the generated plans in experiments with human subjects across several domains and found them significantly more suspicious than the baselines, without being obviously adversarial. Our generated Loitering behaviour was deemed suspicious for the longest across all domains.
@inproceedings{prbgi26ws,
title = {Generation of Suspicious Behavior},
author = {Peta Masters and Daniel Gallagher and Gal A. Kaminka and Mor Vered},
booktitle = {Proceedings of the {ICAPS} Workshop on Planning and Reasoning About Beliefs, Goals, and Intentions ({PR-BGI})},
year = {2026},
abstract = {
Patterns of behaviour are regarded as suspicious when they deviate from what is normal or expected for a given domain, where the motive is unknown and suggestive of risk. Recognising suspicious behaviour is valuable because it enables effective threat detection and response but in order to train machine learning models and security personnel to identify such behaviour, we must first define it and provide examples from which they can learn. Suspicion is a human construct, a property of the observer rather than of the behaviour itself, therefore we base our approach on goal recognition, which focuses intrinsically on the observer's perspective. We define and demonstrate several types of suspicious behaviour, synthesising plans for \emph{loitering}, \emph{obfuscation}, and \emph{deception}. These reflect the behaviour of an agent that believes it may be being observed, and acts to conceal its intent.
We evaluated the generated plans in experiments with human subjects across several domains and found them significantly more suspicious than the baselines, without being obviously adversarial. Our generated Loitering behaviour was deemed suspicious for the longest across all domains.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Tue Jun 23, 2026 03:09:46