User Verification Based on Mouse Dynamics: a Comparison of Public Data Sets

mouse_dynamics_balabit_chaoshen_dfl

Context

Code repository of paper: 13th International Symposium on Applied Computational Intelligence and Informatics (SACI), IEEE pp. 143-147, 2019.

Used data sets

• Balabit - Mouse Dynamics Challenge (10 users)
• Chaoshen - Chao Shen's data set (28 users)
• DFL - Our DFL data set (21 users)

Evaluation

Feature extraction

Raw data were segmented into mouse actions then 39 features were extracted from each mouse action. For details see

• Our paper Intrusion Detection Using Mouse Dynamics
• Resources

Performance measures

Performances are reported using ROC Area Under Curve (AUC).

Evaluation protocol

A binary classifier (Random forest, 100 trees) was trained for each users using positive and negative data. In the case of positive data, the chronologically first 2/3 of the data was used for training and the remaining 1/3 of the data was used for evaluation. Negative data were selected from the other users (#positive samples = #negative samples).

• First Scenario: user identity predictions using a single mouse action
• Second Scenario: user identity predictions using a sequence of consecutive actions (sliding window, overlap between consecutive windows was 90%).

Software: Python 3, scikit-learn 0.19.1

Usage

Example: Evaluate the Balabit data set using the first 500 actions/user and 10 actions for user identity predictions.

1. Please set the followings in util/settings.py

• CURRENT_DATASET = DATASET.BALABIT
• DATASET_USAGE = DATASET_AMOUNT.FIRST1000
• NUM_TRAINING_SAMPLES = 500
• NUM_ACTIONS = 10

2. Run evaluation

• python main.py