|
Algorithm | Ref. | Feature processing method | The experimental results |
|
Rule-based algorithm | [5] | Eight features such as User-Agent and http-accept are selected. | The algorithm requires some prerequisites, there is a 65% probability that the algorithm will start, and the accuracy rate after startup will be 99.1%. |
Clustering algorithm | [45] | Four feature weight distribution methods were selected: (1) average weight, (2) entropy as a weight, (3) disagreement decay as a weight, and (4) consider both entropy and disagreement decay. | It is best in the case of feature processing scheme 4, with an accuracy rate of 99.98%, the precision is 93%, and the recall is 87%. |
Algorithm based on feature similarity | [46] | Measured by Levenshtein distance of Pluginlist in the browser, the author selected different thresholds and access intervals for the experiment. | When the threshold is set to 60, the accuracy reaches 97.94%. When the threshold is 53, the accuracy is still 97.57 when the access interval exceeds four weeks |
Random forest | [8] | Eight features were selected according to their influence on the classification results. Random forest selects 10 trees and 3 features. | The best ownership of the article is 0.985, which means that for long-term tracking of each browser, only 1.5% match errors. |
LSTM | [47] | In this paper, fingerprint features are transformed into one-dimensional vectors, and the latest three fingerprints are input into a group each time. | The best accuracy of the training set is 92.4%, and the best accuracy of the test set is 93.3%. |
Bi-RNN | [49] | The authors split the UserAgent attributes and then weighted the sum. The Canvas element is CRC replaced. | The best F1-score of this method reached 99.25%. |
|