Using one class SVM to counter intelligent attacks against an SPRT defense mechanism

摘要

A widely used defense mechanism in current wireless sensor networks is based on using a Sequential Probability Ratio Test (SPRT). This test does not require a fixed sample size, and this reduces the number of communications required and the battery consumption, both of which are of capital interest in such networks. However, SPRT assumes that the distributions under test do not change, and this assumption needs not hold for current attackers.We divide this work in two parts, First, we develop an optimal attack strategy against a Bernoulli SPRT mechanism which is used in many defense mechanisms in current literature. The control law for such an attacker turns out to be easy to implement and very effective, thus posing a significant threat for the defense mechanisms that use such SPRT. Second, we make use of One Class Supporting Vector Machines to obtain a modified SPRT test that is able to detect, not only such an attacker, but potentially any other attack mechanism that has not a similar spectrum to the expected signal from normal sensors. Our work is validated via simulations, showing that the attacker we propose is a real threat to SPRT mechanisms, but also, that our proposed defense mechanism can efficiently cope with such an attacker. (C) 2019 Elsevier B.V. All rights reserved.