Robustness Of URL Online Learning

The Uniform Resource Locator(URL)benefits ordinary users with sharing and accessingi nformati on,but also creates an opportunity for adversari es.A dversari es may create fake URLsto attract visits and steal private information and assets from users.Researchers apply onlinelearning algorithms to solve the malicious URL detection problem.However,online learningin the adversarial environment would be sensitive to causative attack because of its dynamiclearning mode.The robustness of online classifiers may be challenged by adversarial attacks.This paper studies the influence of causative attack on malicious URL online detection,andpresent defense solutions for online learning.The major contri buti ons are as follows:1)The first issue is what kind of learning flaw exists in online classifier training.Onlinelearning under adversarial environment is a game between attacker and defender.We need tostudy the learning flaw in online classifier under adversarial environments first,so that it willhel p defender to i mprove the robustness of onl i ne models.F or mal i ci ous U R L onl i ne detecti on,we suppose that attackers can add spuri ous features to mal i ci ous sampl es to achi eve thei r attackpurpose and present confusion and omission attack strategy.This thesis compares theperformance of four online learning algorithms under attacks through experiments and alsocompares those attack strategies under different settings of attack purpose,knowledge andability.Experimental results show that attack strategies confuse online classifiers and decreasetheir prediction accuracies.2)T he second issue is how to i mprove the robustness of onl i ne I earni ng agai nst adversarialattacks.Defense against adversarial attack is the final objective of adversarial learning,so weneed to find solutions to make sure of the availability of online learning.To deal with aboveattack strategies,this paper presents feature selection and sample replacement strategies.Wedelay the training process to get some testing data and then evaluate the current update step.This article compares the performance of four online learning algorithms using defensestrategies through experiments,and also compares the defense effects under different attackpercentage and observati on step setti ngs.Finally,welist the running time of defense strategies.Experimental results show that our defense methods finish calculations in linear time and are immune to some adversarial attacks.