Counter intrusion software : malware detection using structural and behavioural features and machine learning

Over the past twenty-five years malicious software has evolved from a minor annoyance to a major security threat. Authors of malicious software are now more likely to be organised criminals than bored teenagers, and modern malicious software is more likely to be aimed at stealing data (and hence money) than trashing data. The arms race between malware authors and manufacturers of anti-malware software continues apace, but despite this, the majority of anti-malware solutions still rely on relatively old technology such as signature scanning, which works well enough in the majority of cases but which has long been known to be ineffective if signatures are not updated regularly. The need for regular updating means there is often a critical window---between the publication of a flaw exploitable by malware and the distribution of the appropriate counter measures or signature. At this point a user system is open to attack by hitherto unseen malware. The object of this thesis is to determine if it is practical to use machine learning techniques to abstract generic structural or behavioural features of malware which can then be used to recognise hitherto unseen examples. Although a sizeable amount of research has been done on various ways in which malware detection might be automated, most of the proposed methods are burdened by excessive complexity. This thesis looks specifically at the possibility of using learning systems to classify software as malicious or nonmalicious based on easily-collectable structural or behavioural data. On the basis of the experimental results presented herein it may be concluded that classification based on such structural data is certainly possible, and on behavioural data is at least feasible.