Security of ML Models

Inspiration

Inspiration: The project was inspired by the growing need for secure, real-time face authentication in a browser environment. The challenge was to ensure both model security and efficiency while catering to a population-scale system like UIDAI's authentication.

What I Learned: I deepened my understanding of model obfuscation, encryption, and how crucial security measures are in edge AI deployments. The project also enhanced my knowledge of TensorFlow Lite optimization, browser caching mechanisms, and containerized cloud deployment.

How I Built It: I started by training a TensorFlow model, converting it into a TFLite model, and optimizing its size to ensure minimal latency. I implemented obfuscation techniques like renaming, parameter encapsulation, and dummy layer injection. The model was then encrypted and deployed through a scalable backend in AWS, using Docker for containerization. Browser caching and secure inference were incorporated for seamless, real-time authentication.

Challenges Faced: The major challenges were achieving model security without increasing file size significantly, integrating real-time inference within a browser, and handling high transaction volumes without compromising performance. Additionally, ensuring the model update process didn’t disrupt the system’s efficiency required innovative obfuscation and encryption cycles.

What it does: It delivers a secure ML model for face authentication by encrypting, obfuscating, and optimizing the model. It ensures secure transactions with minimal size impact, caching the model in the browser for reuse.

How we built it: We trained a TensorFlow model, converted it to TFLite, applied obfuscation methods, encrypted it, and deployed it via Docker on AWS for real-time use. The model is cached and reloaded when needed.

Challenges we ran into: Balancing security without increasing model size and ensuring efficient inference in the browser were key challenges, along with scalable deployment.

Accomplishments that we're proud of: Successfully creating a secure, scalable, and optimized solution that can handle high transaction volumes while maintaining real-time authentication accuracy.

What we learned: Deep insights into obfuscation, encryption techniques, cloud deployment, and secure edge AI applications in browser environments.

What’s next for Security of ML Models: Further enhancements could include advanced encryption algorithms, more efficient model update systems, and expanding the solution to other real-time browser-based AI applications.