Login Signup Classifier - ML Model

This repository contains the data, notebooks and the model output that we used in our paper The Double-Edged Sword: Identifying Authentication Pages and their Fingerprinting Behavior.

The paper introduces a novel machine learning-based approach for automated identification of authentication pages (i.e. login and signup pages) and measures the prevalence of fingerprinting scripts on those pages. For a more detailed overview visit the main repository. Additionally, to explore the browser extension utilizing this ML model for identifying login and signup pages, you can visit the corresponding repository.

ML Pipeline

Model Training

We used the TensorFlow framework to train a multi-class classifier. For each visited page, we generate an 88-dimensional feature vector and a label, which we fed to a neural network with two hidden layers containing 8 and 16 units. The output layer is mapped to the three classes, i.e., login, sign-up, and neither. We trained the model for 200 epochs using the cross-entropy loss function. The notebook along with the required data that we used to train our model made publicly available with this repository.

Data

The data from the crawl (performed on August'23) will be made available soon. The auxiliary data utilized in our analysis and model training comprises the following:

1. 2023_08_25_top_100k_crawl_res_df.csv: Contains information regarding the pages visited by our crawler, indicating whether they are login or signup pages.
2. 2023_08_25_top_100k_fps.csv: Provides details about fingerprinting scripts found on each visited page, including fingerprinting type(canvas, font etc), domain, etc.
3. classifiers_comparison_random_sample.csv: A randomly selected sample of 1000 site URLs used to evaluate the performance of different login/signup page classifiers (including our ML model, Fathom, Autofill, and previous heuristics).
4. crux_fqdn_top_100k_202304_with_rank.csv: Includes the rank of the URLs visited during the analysis.
5. inner_links.csv: Records the inner links visited by each homepage examined.
6. signals.csv: Contains login and signup-related page signals used in model training.

Reference

@article{
    author    = {Asuman Senol, Alisha Ukani, Dylan Cutler and Igor Bilogrevic},
    title     = {{The Double Edged Sword: Identifying Authentication Pages and their Fingerprinting Behavior}},
    booktitle = {Proceedings of The Web Conference 2024},
    year      = 2024,
    month     = May
}