WebGraph

Update (30/07/2024): We are currently resolving some issues related to OpenWPM compatibility with Firefox for WebGraph crawls. We will release fixes soon. In the meantime, a trained model of WebGraph on 20k sites can be found here: https://drive.google.com/drive/folders/1IivW179R5YHYCqOMFH77LxyZ0aJixqFV?usp=drive_link . This model was trained on April 6-8 2023.

Artifact release for the paper "WebGraph: Capturing Advertising and Tracking Information Flows for Robust Blocking", published at USENIX Security 2022.

---

Requirements

Installation

This project has been run and tested on Ubuntu 18.04

First, make sure you have python3, binutils, pip, gcc and g++ installed. Otherwise run the following command

apt-get install binutils python3-dev python3-pip gcc g++

To run all tasks (Graph building, Feature extraction or Classification) on WebGraph, the crawl data used is collected using a custom version of OpenWPM. Follow the instructions here to setup OpenWPM in your environment.

After OpenWPM is installed, if you haven't done it yet, activate the conda environment:

conda activate openwpm

go into <project-directory>/code in the project folder and install the python libraries in requirements.txt:

pip install -r requirements.txt

Preparing Crawl Data

To generate the crawl data needed for the pipeline, you need to run a crawl using the installed OpenWPM tool. To run a crawl, first update the script demo.py to read in the list of sites that you want to visit. Then, run demo.py.

After you run the demo, a datadir folder will be created in your demo directory. Inside the folder, you will find two database files to be used in our pipeline: crawl-db.sqlite and content.ldb

Pipelines

The codebase consists of two pipelines: WebGraph and Robustness. We describe each of them below.

WebGraph Pipeline

This pipeline runs the WebGraph system, which is a graph-based Ad and Tracking Services (ATS) detection system. WebGraph takes in crawl data, builds graph representations of sites, extracts features and labels from these representations, and trains a machine learning model.

With the WebGraph code, we present two tasks that you can run:

1. Graph Preprocessing and Feature building
2. Classification (training and testing)

1. Graph preprocessing and Feature Building

In this task, WebGraph constructs the dataset for classification by:

• taking your sqlite and leveldb database files to construct a graph representation of each crawl as explained in the paper and export it in a tabular format to a graph.csv file and features.csv file
• applying the rules from public filterlists to label the nodes in each graph and export it in a tabular format to a labelled.csv file

To run this task, run the following script:

python <project-directory>/code/run.py --input-db <location-to-datadir>/datadir/crawl-db.sqlite --ldb <location-to-datadir>/datadir/content.ldb --mode webgraph

All additional arguments accepted by this command:

• --input-db: the path to the .sqlite file generated by the crawl
• --ldb: the path to the .ldb file generated by the crawl
• --features: the path to the .yaml feature categories list. A default features.yaml is used if unspecified.
• --filters: the path to the directory to save the filter lists in. A default filterlists folder will be created if unspecified.
• --out: the path to the directory of the output .csv files.
• --mode: the system to run (webgraph or adgraph).

Note: With the --mode argument, you can also run AdGraph (we evaluate AdGraph in Section 3 of the paper).

2. Classification

The classification takes in the output from Step 1 (features and labels), and performs cross validation on the data. To run this task, run the following script:

python <project-directory>/code/classification/classify.py --features features.csv --labels labels.csv --out results --save False --probability False --interpret False

Arguments of this command:

• --features: the path to the features.csv file
• --labels: the path to the labels.csv file
• --out: the path to the directory of the output files
• --save: Whether to save the trained model.
• --probability: Whether to save prediction probabilities.
• --interpret: Whether to run interpretation on results.

---

Robustness Pipeline

This pipeline runs the robustness experiments performed in the paper. There are two types of robustness experiments: content and structure mutations. All the code and READMEs associated with these experiments are in the robustness folder.

Data Schema

The output of the WebGraph pipeline is three files: graph.csv, features.csv, labelled.csv.

Graph

These are the columns present in the graph output under graph.csv

Column	Applies to	Description
visit_id	All	the visit id of the crawl
name	All	the name of the node or edge
graph_attr	All	Node or Edge
top_level_url	All	The top level URL (page being visited)
attr	All	additional attributes of nodes and edges
domain	All	The parent domain of nodes or edges
top_level_domain	All	Top level domain (domain of page being visited)
type	Node	The type of node `Document
document_url	Node	Context of a script's execution.
setter	Node	The name of the node that sets a storage node.
setting_time_stamp	Node	Time stamp of storage node setting.
setter_domain	Node	Domain of the node that sets a storage node.
party	Node	The partiness of a node either first or third or N/A
src	Edge	The source node name of the edge
dst	Edge	The destination node name of the edge
reqattr	Edge	HTTP request headers
respattr	Edge	HTTP response headers
response_status	Edge	HTTP response status
content_hash	Edge	Content hash if logged by OpenWPM
post_body	Edge	POST response body hash
post_body_raw	Edge	POST response body raw

Features

The features in features.csv used are described in features.yaml

Labels

Nodes labeled by either True or False if they are blocked by filter lists or not. These are the columns present in the labelled.csv file.

Column	Description
visit_id	The visit id of the crawl
top_level_url	The top level URL (page being visited)
name	The name of the node
label	The label of the node

---

Code Organization

The WebGraph pipeline is in the code folder. The Robustness pipeline is in the robustness folder.

Paper

WebGraph: Capturing Advertising and Tracking Information Flows for Robust Blocking Sandra Siby, Umar Iqbal, Steven Englehardt, Zubair Shafiq, Carmela Troncoso USENIX Security Symposium (USENIX), 2022

Abstract -- Users rely on ad and tracker blocking tools to protect their privacy. Unfortunately, existing ad and tracker blocking tools are susceptible to mutable advertising and tracking content. In this paper, we first demonstrate that a state-of-the-art ad and tracker blocker, AdGraph, is susceptible to such adversarial evasion techniques that are currently deployed on the web. Second, we introduce WebGraph, the first ML-based ad and tracker blocker that detects ads and trackers based on their action rather than their content. By featurizing the actions that are fundamental to advertising and tracking information flows – e.g., storing an identifier in the browser or sharing an identifier with another tracker – WebGraph performs nearly as well as prior approaches, but is significantly more robust to adversarial evasions. In particular, we show that WebGraph achieves comparable accuracy to AdGraph, while significantly decreasing the success rate of an adversary from near-perfect for AdGraph to around 8% for WebGraph. Finally, we show that WebGraph remains robust to sophisticated adversaries that use adversarial evasion techniques beyond those currently deployed on the web.

The full paper can be found here.

Citation

If you use the code/data in your research, please cite our work as follows:

@inproceedings{Siby22WebGraph,
  title     = {WebGraph: Capturing Advertising and Tracking Information Flows for Robust Blocking},
  author    = {Sandra Siby, Umar Iqbal, Steven Englehardt, Zubair Shafiq, Carmela Troncoso},
  booktitle = {USENIX Security Symposium (USENIX)},
  year      = {2022}
}

Contact

In case of questions, please get in touch with Sandra Siby.

Acknowledgements

Thanks to Laurent Girod and Saiid El Hajj Chehade for helping test and improve the code.