ghadiaravi13 / Mouse-Dynamics

This repository contains the code for usage of mouse dynamics in imposter detection.

Mouse-Dynamics for imposter detection

Machine Intelligence and Expert Systems Term Project,

Autumn Semester, 2018-19,

Department of Electronics and Electrical Communication Engg,

IIT Kharagpur

Required python packages:-

1.Numpy

2.Sklearn

3.pandas

Execution of code:-

• Open terminal or ide and run the main.py file
• Note: The text files containing the data and python files should be in same folder

Steps followed:-

Dataset Collection:-

• Reference:- Folder named '\data'
• Collected using the mouse.jar application.
• Continuous data was collected over a period of time by various users

Extracting Dataset:-

• Reference:- extractor.py
• Pre-processes the raw data and transform it to contain hold time and latencies for all combination of keystrokes,
• The basics of mouse movement: X-coordinate, Y-coordinate, Theta value etc are extracted.

Naive- Bayes Classifier:-

• Reference:- main.py
• Data obtained from extractor.py is used as input.
• Data for each user is assigned a particular class value (0,1,2,..).
• train-test split is done separately for each class to ensure train and test set contain appropriate proportions of each class
• Whole data is then merged, while maintaining the train-test split.
• Gaussian Naive-Bayes Model is implemented on the split X_train, X_test, y_train, y_test
• To validate the stability of the model, five fold cross validation is used.

Execution of code:-

• Open terminal or ide and run the main.py file
• Note: The text files containing the data and python files should be in same folder

Results:-

OBTAINED ACCURACY

1.When number of classes(users) is 5:-

       - GaussianNB:- 49.49833% 
       - GaussianNB with 5-fold Cross validation:- 15.798%

2.When number of classes(users) is 8:

       - GaussianNB:-  83.49355%
       - GaussianNB with 5-fold Cross validation:- 67.39%

3.When number of classes(users) is 9:

        - GaussianNB:-  94.75419%
        - GaussianNB with 5-fold Cross validation:- 80.674%

Conclusion:

We can witness a significant increase in the accuracy when the training dataset is increased.