Skip to content

Latest commit

 

History

History
95 lines (66 loc) · 3.16 KB

README.md

File metadata and controls

95 lines (66 loc) · 3.16 KB

Refract: Protect Your Digital Identity

Team Members

  • Ysabel Chen
  • Gen Tamada
  • Emily Tian
  • Jesus Velarde
  • Christy Yu

Overview

alt text

Refract is a cutting-edge tool designed to cloak photos of individuals, safeguarding them against:

  • Unauthorized training for deepfakes.
  • Facial recognition systems.
  • Other harmful forms of identity theft.

Our approach involves subtly altering images at the feature level to disrupt the ability of image-processing models to identify individuals. Refract ensures these changes are:

  • Human imperceptible: Visual differences are nearly undetectable.
  • Highly effective: Alters feature-identifying capabilities of image-processing models.

With Refract, individuals can confidently upload photos to social media, mitigating the risks associated with facial recognition technologies, identity theft, and deepfake generators.

How It Works

alt text

  1. Target Image Selection: Refract identifies a target image similar to the input photo.
  2. Feature Space Alteration:
    • Introduces subtle changes to the original image to match the target photo’s feature space.
    • Alternatively, modifies the input photo to maximize changes in its own feature space.
  3. Visual Consistency: The final image retains strong visual similarity to the original photo while disrupting model-identifying capabilities.

Features

  • Multiple Iterations: Experimented with five model iterations, refining the process for optimal results.
  • User Customization: A demo frontend includes sliders for adjusting:
    • Cloak strength.
    • Other parameters for balancing safety and visual distortion.
  • Alternative Methods: Explored approaches including:
    • Altering the feature space of the source image itself.
    • Mimicking the feature space of a similar-looking target image.

Technology Stack

Backend

alt text

  • AWS Infrastructure:
    • QueueJob for task scheduling.
    • S3 buckets for secure image storage.
    • EC2 instances for model hosting and computation.
  • Model Framework:
    • Built using PyTorch for deep learning and adversarial techniques.

Frontend

  • React Web Application:
    • Hosted on Vercel for seamless deployment.
    • Interactive sliders for user customization.

Key Achievements

  • Successfully demonstrated the ability to alter photos on the feature level while maintaining visual fidelity.
  • Created a scalable and user-friendly tool for protecting identity online.
  • Developed a flexible and robust system that integrates modern cloud infrastructure with advanced machine learning.

How to Use

  1. Upload a photo to the demo interface.
  2. Adjust the sliders to set cloak strength and other parameters.
  3. Download your cloaked image and share it confidently!

Why Refract Matters

Refract addresses growing concerns about privacy and security in the digital age. By disrupting the feature-identifying capabilities of image-processing systems, we empower individuals to take control of their digital identity.

For more details, visit our project or contact the team!

Demo: Refract