Intrusion systems have been the subject of considerable research during the past 33 years, since the original work of Anderson. Much has been published attempting to improve their performance using advanced data processing techniques including neural nets, statistical pattern recognition and genetic algorithms. Whilst some significant improvements have been achieved they are often the result of assumptions that are difficult to justify and comparing performance between different research groups is difficult. The thesis develops a new approach to defining performance focussed on comparing intrusion systems and technologies. A new taxonomy is proposed in which the type of output and the data scale over which an intrusion system operates is used for classification. The inconsistencies and inadequacies of existing definitions of detection are examined and five new intrusion levels are proposed from analogy with other detection-based technologies. These levels are known as detection, recognition, identification, confirmation and prosecution, each representing an increase in the information output from, and functionality of, the intrusion system. These levels are contrasted over four physical data scales, from application/host through to enterprise networks, introducing and developing the concept of a footprint as a pictorial representation of the scope of an intrusion system. An intrusion is now defined as “an activity that leads to the violation of the security policy of a computer system”. Five different intrusion technologies are illustrated using the footprint with current challenges also shown to stimulate further research. Integrity in the presence of mixed trust data streams at the highest intrusion level is identified as particularly challenging. Two metrics new to intrusion systems are defined to quantify performance and further aid comparison. Sensitivity is introduced to define basic detectability of an attack in terms of a single parameter, rather than the usual four currently in use. Selectivity is used to describe the ability of an intrusion system to discriminate between attack types. These metrics are quantified experimentally for network intrusion using the DARPA 1999 dataset and SNORT. Only nine of the 58 attack types present were detected with sensitivities in excess of 12dB indicating that detection performance of the attack types present in this dataset remains a challenge. The measured selectivity was also poor indicting that only three of the attack types could be confidently distinguished. The highest value of selectivity was 3.52, significantly lower than the theoretical limit of 5.83 for the evaluated system. Options for improving selectivity and sensitivity through additional measurements are examined.

Document Type


Publication Date