Label flipping attacks against Naive Bayes on spam filtering systems

Abstract

Label flipping attack is a poisoning attack that flips the labels of training samples to reduce the classification performance of the model. Robustness is used to measure the applicability of machine learning algorithms to adversarial attack. Naive Bayes (NB) algorithm is a anti-noise and robust machine learning technique. It shows good robustness when dealing with issues such as document classification and spam filtering. Here we propose two novel label flipping attacks to evaluate the robustness of NB under label noise. For the three datasets of Spambase, TREC 2006c and TREC 2007 in the spam classification domain, our attack goal is to increase the false negative rate of NB under the influence of label noise without affecting normal mail classification. Our evaluation shows that at a noise level of 20%, the false negative rate of Spambase and TREC 2006c has increased by about 20%, and the test error of the TREC 2007 dataset has increased to nearly 30%. We compared the classification accuracy of five classic machine learning algorithms (random forest(RF), support vector machine(SVM), decision tree(DT), logistic regression(LR), and NB) and two deep learning models(AlexNet, LeNet) under the proposed label flipping attacks. The experimental results show that two label noises are suitable for various classification models and effectively reduce the accuracy of the models.