Deep Learning–based Watermarking Systems

- MTech Thesis

National Institute of Technology Patna, 2022
📄 PDF: Technical Report

Themes: Deep Learning, Image Privacy, Watermarking, CNNs, Autoencoder Architectures

Summary: This is my MTech research (done over 1 year) that extends image watermarking investigations using deep learning. It includes algorithmic design and robustness testing, forming the basis of the SPIE journal publication.

Overview: Developed a deep learning model that hides one image inside another—like embedding an image watermark (from the MNIST dataset) invisibly into a completely different photo. 

The model can later extract the hidden image [watermark] accurately, even if the outer image is compressed or slightly distorted. The robustness has been tested across different attacks. This technique offers a powerful way to protect digital image ownership using AI.

⚙️ Technical Highlights:
1. Neural Networks Used: Convolutional Neural Networks (CNNs), Autoencoders
2. Architecture: Two-part system (embedding + extraction models)
3. Tech Stack: Python, TensorFlow
4. Key Techniques: Feature map fusion, transposed convolutions, ReLU activation
5. Performance: High invisibility and robustness; outperformed other methods under multiple distortion attacks