Pardah

Inspiration

Generations of Muslims have experienced a multitude of trials, and with the rise of the digital era, today’s generation faces a new one: the permanence of our online presence.

When two of our teammates began their hijab journey in the summer of 2025, they encountered an obstacle: despite their commitment to modesty, they kept unexpectedly discovering old images of themselves without hijab circulating throughout the web—Instagram, Facebook, Google Images, and random websites they didn’t even know existed.

The existence of these images made some new hijabis reluctant to start covering their hair. Their beautiful choice to embrace the hijab was being undermined by their digital past. Existing privacy tools weren’t designed to address this specific need, and the hours spent manually untagging themselves from Instagram posts, reporting Google Images, and filling out takedown requests proved exhausting and inefficient.

Pardah emerged from this gap—a solution created by Muslim women who understand and have experienced the struggle of maintaining modesty when the internet tries to undermine it.

We believe Muslim women should not have to choose between their faith and their digital footprint. We built Pardah because every Muslim woman deserves the right to control her image, both in person and online, and the freedom to embrace modesty without compromise.

What it does

Pardah was designed for a simple, specific task and with empathy and accessibility in mind. The landing page prompts the user to upload an image of themselves wearing hijab and enter their full name. Upon pressing submit, the program will load and the user will see a stack of a few dozen images of themselves found on the internet. The user will be able to easily move back and forth between these images through the forward and back arrows on the right and left sides of the stack. For each image, the user will see two button options underneath the image:

“Remove”: Upon clicking “Remove”, the program will load a pop-up card at the bottom of the page containing the photo and a mention of the source where the image was found. Suppose there is an opportunity for the user to remove the image directly. In that case, a “Remove” button will be available, directing the user to a link where they can remove the image. If the image cannot be directly removed, a
"Send email to ___” button will be available, allowing the user to send an automated takedown request email to the owner of the site where the image was found.
“This isn’t me”: If an image found was not the user's, the user can click this button, and the image will be removed from the stack and replaced with a placeholder card.

How we built it

Frontend Stack

React (v19.2): Builds the UI with reusable components and hooks
Axios: Sends and receives data from the Flask backend
React Hooks: useState for managing component state
CSS3: Custom responsive styling

Backend Stack

Logistics

Flask: Python web framework for creating REST API endpoints
Flask-CORS: Enables cross-origin requests from React frontend
Python-dotenv: Environment variable management

Libraries

DeepFace: AI library for facial recognition using FaceNet512 model
OpenCV: Image processing and computer vision operations
Pillow (PIL): Image manipulation and format conversion
NumPy: Mathematical operations on face embeddings (512-dimensional vectors)
Bing Image Downloader: Web scraping tool to search and download images
Requests: HTTP library for downloading images from URLs
Werkzeug: Secure filename handling and file upload management
Hashlib: SHA256 image hashing for reporting

Class Structure

search_images.py: Crawls web to find images of subject given keyword
face_recognition.py: Handles face embedding and comparison tasks
report_generator.py: Connects users with tools for reporting and takedown requests

Basic Flow

User agrees to image being processed locally → User uploads photo + enters name → React frontend packages inputs & sends to backend → search_images.py runs → crawls Bing for matching images → downloads results → face_recognition.py runs → extracts face embeddings from user photo + search results → compares embeddings → calculates Euclidean distances + similarity scores → React frontend receives results → displays all matched images with similarity scores → User reviews images → confirms true matches → When user wants to remove an image, report_generator.py runs → Bing removal email template → Step-by-step removal plan

Challenges we ran into

Running backend on the correct server: Lining up the frontend image upload with the backend search; uploads were not working
Search Result Quality: Scraping Bing, Instagram, and Facebook, trying to gather as many results as possible. Website blockers and image downloading issues
Face Detection Accuracy: Ensuring the imageset returned the correct face and photos so users did not have to look through photos that were not of them
Match Confidence: Tested different similarity thresholds
Buggy Interface: Textbox glitches

Accomplishments that we're proud of

Getting the Web Crawling to work: Returning actual results and gathering image data from multiple platforms
Facial Recognition Accuracy: Achieved 97.53% accuracy!
Developer Tools: Learned how to use existing tools and online resources to optimize coding
User Experience Design: Created an aesthetically pleasing interface for our audience, with color scheming and fun fonts

What we learned

Data and User Privacy: The importance of protecting user information and verification processes
Pre-existing Tools: Took advantage of existing scrapers and APIs for a more effective product
UI Flow: Followed a “less is more” design philosophy, as we are simplifying a typically complex reporting process

What's next for Pardah

We’re eager for the further development of Pardah and are excited to see where it takes us! Looking forward, we have three main advancements we would like to implement:

Form Autofill: Automatically filling out reporting forms for users (e.g., via a Chrome extension) on various platforms
Faster Crawling: Return images faster
Identity Verification: Through Duo or similar ID-based verification processes
- Examples: Meta has age verification, LinkedIn has ID-photo verification
- This ensures only the hijabi can manage the reporting of her images and prevents harmful misuse of the tool.
Search Engine Integration: Implementation into Google and other search engines
- Access to larger datasets and more searched images