Improving the Resilience of Machine Learning Models to Adversarial Attacks for Cross-Site Scripting Detection

Abstract

This paper proposes an approach to increase the resilience of machine and deep learning algorithms to adversarial attacks. The paper considers various ways to conduct cross-site scripting attacks. A number of studies on the use of machine and deep learning (ML/DL) algorithms for detecting cross-site scripting attacks are analyzed. A general algorithm for conducting adversarial attacks on ML/DL algorithms is described. A scenario for an attack on a detector model operating in the "black box" mode is considered. The method for generating adversarial examples is reinforcement learning. MLP, CNN, and LSTM were selected as the models to be attacked. To generate adversarial attacks, the Proximal Policy Optimization (PPO) algorithm is used, which provides more stable training of the attacking model and has less sensitivity to hyperparameters, while maintaining sufficient performance. The solution based on the PPO algorithm uses selected mutations checked for syntactic correctness, which are then used to replenish the dataset when retraining the detector model. The experimental results demonstrate the efficiency and applicability of the proposed solution. Standard quality metrics were used. Despite the initially high F1-measure value, all models missed 98% or more of the adversarial examples, demonstrating complete instability to the adversarial attack. The authors implemented twelve mutations of XSS attacks, four of which showed the greatest efficiency. In addition to generating adversarial examples, logging of the obtained examples was implemented to form a dataset for additional training, as well as collecting statistics on successful and unsuccessful mutations.