Quantitative analysis of the release order of defensive mechanisms

Abstract

Dependency on information technology (IT) and computer and information security (CIS) has become a critical concern for many organizations. This concern has essentially centred on protecting secrecy, confidentiality, integrity and availability of information. To overcome this concern, defensive mechanisms, which encompass a variety of services and protections, have been proposed to protect system resources from misuse. Most of these defensive mechanisms, such as CAPTCHAs and spam filters, rely in the first instance on a single algorithm as a defensive mechanism. Attackers would eventually break each mechanism. So, each algorithm would ultimately become useless and the system no longer protected. Although this broken algorithm will be replaced by a new algorithm, no one shed light on a set of algorithms as a defensive mechanism. This thesis looks at a set of algorithms as a holistic defensive mechanism. Our hypothesis is that the order in which a set of defensive algorithms is released has a significant impact on the time taken by attackers to break the combined set of algorithms. The rationale behind this hypothesis is that attackers learn from their attempts, and that the release schedule of defensive mechanisms can be adjusted so as to impair the learning process. To demonstrate the correctness of our hypothesis, an experimental study involving forty participants was conducted to evaluate the effect of algorithms’ order on the time taken to break them. In addition, this experiment explores how the learning process of attackers could be observed. The results showed that the order in which algorithms are released has a statistically significant impact on the time attackers take to break all algorithms. Based on these results, a model has been constructed using Stochastic Petri Nets, which facilitate theoretical analysis of the release order of a set of algorithms approach. Moreover, a tailored optimization algorithm is proposed using a Markov Decision Process model in order to obtain efficiently the optimal release strategy for any given model by maximizing the time taken to break a set of algorithms. As our hypothesis is based on the learning acquisition ability of attackers while interacting with the system, the Attacker Learning Curve (ALC) concept is developed. Based on empirical results of the ALC, an attack strategy detection approach is introduced and evaluated, which has achieved a detection success rate higher than 70%. The empirical findings in this detection approach provide a new understanding of not only how to detect the attack strategy used, but also how to track the attack strategy through the probabilities of classifying results that may provide an advantage for optimising the release order of defensive mechanisms.