| The traditional online user recognition method based on cookies identifies the user through the identification information stored in the client.However,when users delete cookie files,this method will be greatly limited.The user recognition method based on browser fingerprints uses the browser fingerprint information to identify users,because the acquisition of browser fingerprint information does not need the participation of users and does not need to store information in the client,so this method has a good application prospect.However,due to the evolution of browser fingerprint over time,a difficult problem of user recognition method based on browser fingerprint is how to identify users according to the dynamic changes of browser fingerprint.This paper focuses on the research of browser fingerprint user recognition method based on Long Short-Term Memory(LSTM)network,and uses LSTM model to learn the evolution law of browser fingerprint over time.The main contents of this paper are as follows:(1)The existing browser fingerprint based recognition methods based on machine learning regard the user recognition as a binary classification problem.The method compares the browser fingerprints one by one in data processing,and finally takes the comparison results as the input vector of the recognition model.This method of data comparison will cause information loss,and when using this model for fingerprint recognition,it is necessary to compare the fingerprint to be identified with the fingerprints in the database one by one,so the recognition efficiency is low.In order to solve the above problems,a browser fingerprint recognition method based on multi-class classification LSTM is proposed.The basic idea is to process the same user’s browser fingerprint data into fingerprint sequence in time order,and use multi-class classification LSTM model to classify them,so as to realize user recognition.In theory,the method can process the fingerprint data directly without the need of classification and contrast conversion,thus avoiding the loss of information;In addition,when testing new samples,the problems of comparison and calculation in traditional fingerprint recognition are avoided,and the recognition efficiency is improved.The experimental results show that the fingerprint recognition method based on multi-class classification LSTM has higher accuracy and faster recognition speed than the binary classification based browser recognition method.(2)In practical application,user recognition based on browser fingerprint belongs to a incremental learning task,that is,the browser fingerprints of users will change with time,and new users will also appear.The model trained with new fingerprint data will be updated continuously,which leads to the decrease of adaptability of the model to the old browser fingerprint data,and catastrophic forgetting problem will occur.In order to solve this problem,based on the existing incremental learning model based on distillation loss and bootstrap sampling method,this paper further improves the above proposed recognition method by incremental training.The basic idea is to train a teacher LSTM model by using historical browser fingerprint sequence data.Then,the bootstrap sampling method is used to select part of the historical data as the representative browser fingerprint data.Finally,the cross distillation loss is used as the loss function,and the representative browser fingerprint data and new browser fingerprint data are used to train the student model.The experimental results show that the incremental fingerprint recognition method based on LSTM can reduce catastrophic forgetting,and the model trained by bootstrap sampling method has better recognition performance. |
