Research Of Adversarial Attack Methods For Deep Time-Series Models

In recent years,with the development of sensor technology,the increase of digitalization,the popularity of Internet,and the rise of big data technology,modern society has generated more and more time series data.Usually,research on time series includes two aspects: time series classification and time series prediction.In the past,people usually use statistics-based methods to solve time series classification and time series prediction tasks.With the continuous development of deep learning techniques,deep learning models are also being used more and more widely to solve time series classification and prediction tasks.However,deep temporal models are not immune to the vulnerability of deep models to adversarial attacks.In this paper,we focus on the adversarial attack methods for different deep seriesmodels.First,for deep series classification models,previous adversarial attack methods often focus on improving the success rate of the attack,while ignoring the invisibility of the adversarial samples,which makes the generated adversarial samples differ from the original samples and are easily recognized by the human eye,and do not meet the invisibility requirement of the adversarial samples.The adversarial samples generated by the new adversarial attack method for deep series classification model proposed in this paper are more difficult to be detected compared with the adversarial samples generated by the previous methods,which have better concealment and better comprehensive performance.Second,for deep series prediction model,previous studies have similarly ignored the invisibility of the adversarial samples.When measuring the performance of the adversarial sample,only the magnitude of the mean square error value between the predicted and true values of the adversarial sample is concerned,and the difference between the adversarial sample and the original sample is not taken into account.The attack method with filters proposed in this paper has a better overall performance compared with previous methods,which generate more hidden adversarial samples while still maintaining a good attack effect.Finally,this paper develops a visualization system for deep series model adversarial attack algorithm,which includes several modules such as front-end,back-end and algorithm library,and can show the attack effect of each attack algorithm in the form of images.