AnonymousAds

Targeted ads without user tracking!

Table of Contents

1. About The Project
   • Built With
2. Methodology
3. Getting Started
   • Prerequisites
   • Installation
4. Usage
5. Acknowledgments

About The Project

AnonymousAds is a web application that serves targeted advertising to users, without compromising user privacy. It uses fully homomorphic encryption (FHE) and noise injection to allow for these ads to be served. For more information, you can watch our team's video. This project is a submission to TikTok TechJam 2024 to address Problem Statement 1: Innovating Privacy.

Built With

• ConcreteML
• Flask
• NTLK
• Docker
• Sklearn
• Torch
• Pandas
• numpy
• requests
• shutil

Methodology

Pre-Requisites

The dev container contains all files and code required to generate pre-requisites.

1. The ML model is fitted on clear training data.
2. The model is compiled and then used to generate server.zip and client.zip in the dev/fhe_directory/ directory

Protocol

Fully Homomorphic Encryption (provided by the Concrete-ML library) allows us to apply functions to encrypted data, for them to be in effect, when decryption takes place later. This allows encryption and decryption to be entirely on the local machine, and yet allow the encrypted data to be processed on an external machine. As a proof of concept, this project uses FHE to serve targeted advertising to end-users, while ensuring the privacy of their search history.

1. The search-engine generates public keys and sends it to the server.
2. The user inputs 5 search queries into the search-engine.
3. The search queries is processed into vector, encrypted and is sent to server.
4. The server takes this input runs the model, and sends the still encrypted predictions back to the search-engine.
5. The search-engine decrypts the predictions and updates its predictions.
6. The search-engine makes a request to the server for ads, adding noise to the request, obfuscating the predictions.
7. The search-engine selects the relevant ads from this set of ads, and serves the targeted ads to the user.

Getting Started

AnonymousAds is built to run on desktops.

To get a local copy of AnonymousAds up and running follow these simple steps.

Prerequisites

Docker desktop is required to run the application.

Download Docker at https://www.docker.com/products/docker-desktop/

We recommend having at least 6GB of Disk Space and 4GB of RAM.

This proof of concept was tested on Windows 11 and Ubuntu. It was also tested with Firefox, Google Chrome, and Microsoft Edge. If you face any issues, please free to let us know.

Installation

1. Clone the repo

   git clone https://github.com/Apzure/AnonymousAds.git

2. Open Docker desktop

3. Enter in the terminal

   docker compose up --build

4. Open up the link http://127.0.0.1:5000 provided in the terminal

Usage

Basic Usage

1. When the opening the link above, you will be greeted by this landing page. Enter your queries into the search bar and hit ENTER or click the search button on the right.

2. Initially, there would be no ads displayed on the page.

3. Every 5 queries, the page would be automatically updated with new ads based on your search results. It would take 5-10 seconds for the ads to be updated. See Queries and Keywords for more information.

4. You can view your past queries by clicking on the history icon on the top left corner. Click on the same icon to hide your history.

5. You view the logs on Docker Desktop for more information.

Queries and Keywords

As you input more queries, our search engine tracks and updates your preferences using a Bayesian update scheme. Initially, the engine makes larger adjustments to its predictions as it gathers more information. Over time, as more queries are entered, the model stabilizes, and the ads displayed become more consistent.

Keyword Usage

To help our machine learning model categorize your queries effectively, please include the following keywords. Your queries can contain punctuation and additional words besides the keywords. The keywords are not case-sensitive, and you can use different words with the same "stem" as the keyword. For more about stemming, see the nltk documentation for more about stemming.

Example: "games" shares the same stem "gam" as "gaming" which is a keyword, so "games" would be picked up by the model.

Sports

sports, fitness, athletics, training, running, gear, gym, exercise, sportswear, outdoors, wellness

Food

food, cuisine, gourmet, organic, recipes, delivery, dining, snacks, cooking, restaurants, healthy

Music

music, concerts, streaming, instruments, bands, festivals, songs, albums, lessons, dj, sound

Gaming

gaming, consoles, accessories, esports, multiplayer, virtual, pc, development, merchandise, livestream, communities

TV

tv, dramas, shows, smart, reviews, theater, cable, series, reality, channels, binge, sports, food, music, gaming, tv

Training the model

As described in the methodology above, the server.zip and client.zip files are used to generate the public keys and stores the model for the server. If you would like to train the model yourself and generate these files, you can follow these steps:

1. Enable the dev container by setting ENABLE_DEV_SERVICE=1 in the .env file in the root of the repo

2. Open Docker desktop

3. Enter in the terminal

   docker compose up --build

4. The dev container should be running on Docker Desktop, you can view the logs and track the progress of the model

5. Once the dev container has exited, access the files generated by the container under dev/fhe_directory/. You can save them in your local machine. You may delete the other two containers after this.

6. Replace the respective zip files in search-engine/fhe and server/fhe

7. Run docker compose up --build again, and the model is now the newly trained model!

Note on Ad Selection

Note that we have chosen to display top 2 ads from 2 different categories. This means that if the user has searches that only focuses on one category e.g. sports, which normally only occurs with low quantites of searches, one of the ads may be relatively untargeted.

Acknowledgments

We used royalty-free stock photos in our application. For more information, see docs/credits.txt in our repo.