Biometric authentication with neural networks

This project is an exploration of how we can use neural networks to try to identify a user by using their mouse usage patterns. It's composed of a web application that reads the mouse input and sends it to a backend where a convolutional neural network is executed against this input. The network is pretrained against a dataset of different known users so that it learns how to tell them apart from one another. The question is if we can use the features or representations generated by the network as a kind of signature so that we can distinguish users or detect anomalous patterns.

Background

Identifying users by verifying credentials like passwords or phone access like in two-factor authentication works well enough in many cases. However, credentials and phones can be stolen and can be used by malicious actors to impersonate real users in order to gain access to restricted areas or information.

In the physical world people are often not only identified by credentials but also by other pieces of information like physical traits or behaviours that are more or less unique to each person and very difficult to duplicate. In the case of web applications we can't see the person but we have access to more information than just credentials, for example:

• Mouse and keyboard input patterns.
• Behaviour patterns, how the user interacts with a given application.
• Information about the computer of the user which can be given away by the browser.

With this information we can try to verify the identity of the user or at least become aware of strange patterns which may alert us of a possible impersonation.

Components

The project is organized in several directories:

predictor/ : Contains the code that executes the model, it runs in a separate process. The backend communicates with this process by using an RPC library called Pyro (PYthon Remote Objects).

training/ : Contains the training code and several jupyter notebooks where the dataset and the model performance is explored. There is also a utility to record mouse input from a local user etc.

webapp/ : Contains a django web application with a backend and a frontend so that users can interact with the model.

Launching

• Install python 3.8
• Configure the webapp/webapp/settings.py file, the ALLOWED_HOSTS variable etc..
• Install python dependencies and run the launch script:

pip install -r requirements.txt
./launch.sh