Keystroke-Biometric Authenticator for Android
What is this repository for?
- An Android application showcasing an authentication mechanism for the next generation which eliminates the use of a fixed passcode/password in favor of a machine learning model which uses various signals to generate a unique fingerprint as the user types in randomly generated numeric codes.
- Version 1.0 - Created by Anirudh Ramanathan (firstname.lastname@example.org) and Ashish Kumar (email@example.com) at GTHack 2015.
- Numeric code based authentication schemes can easily lead to stolen credentials.
- Two-factor authentication is cumbersome.
- We have come up with a model which with further improvements, may change authentication works in future.
Early on, we learned from our experiments that if we plotted the screen coordinates at which users touched various numbers on a numeric keyboard on a touchscreen, one would obtain a Gaussian distribution, such as below. The position variation is shown for various keys on a numeric on-screen keyboard for a single user.
Our idea is to use the following signals to build the unique fingerprint of a user, while asking him/her to train a Gaussian multivariate model, adjusting the weights based on confidence probability of all the variate's distribution during training. and then using that model to authenticate the user after that. A major merit of this method is that one is not constrained to using a single passcode/password to authenticate the user. This additionally also gives an extra layer of security over passcodes, because emulating one's use of an electronic device is rather challenging.
We looked at the following signals to generate a unique fingerprint and model for a particular user:
- Typing heatmap (spatial coordinates of points that the user struck)
- Size of contact with the touchscreen
- Duration of contact with the touchscreen
Our application is implemented using Android because it provided to us, APIs that allowed us to query the aforementioned signals, and because we felt that the variety in Android devices would let us test the efficacy of our handpicked signals for the purpose of authentication.
The main screens of the training application are shown below:
This lets us both train the model and use the model we have, to compare against user-provided random codes to check for various similarities in the typing pattern and authenticate them. The toggle-button at the bottom lets one switch between training and authentication modes. It is generally recommended that one train the model with atleast a hundred samples before attempting to run input against the model. The numeric code in the textbox to the top-right is provided as a guideline. It is a set of randomly selected numbers which the user may/may not choose to use. We recommend its use because it produces a well-distributed pattern that is distributed and is likely to reflect one's natural typing pattern.
After each authentication attempt, the application queries the user and records whether it made the right decision. We would like to note that at this point, upon a correct prediction, the application would provide feedback and update the training set. However, we chose not to implement the feedback loop in order to retain the ability to reason about our model in an easier fashion due to time constraints.
The statistics we obtained are shown in the second screen-capture above. This screen can be obtained by clicking on "metrics" from the main application screen. We obtained up to 70% accuracy with our model.
In future, we plan on running further experiments and obtain better accuracy, in order to make it actually feasible to use the application as a replacement for other authentication mechanisms. For e.g. Naive Bayes based probabilistic modelling might improve a lot given the almost independence between consecutive grams. The results we obtained with our model make us very optimistic about the possibilities ahead.
How do I get set up?
- Can be directly imported into Android Studio.
- Dependencies: Included
- Deployment instructions: Android device with Android 5.x or higher recommended.