Are Classification Deep Neural Networks Good for Blind Image Watermarking?

Area of Science:

• Computer Vision
• Digital Image Processing
• Machine Learning

Background:

• Traditional image watermarking relies on Discrete Wavelet Transform or Discrete Cosine Transform, requiring precise synchronization.
• These classical methods often necessitate complex registration mechanisms for embedding and detection accuracy.
• Existing techniques face challenges with synchronization and robustness against image manipulations.

Purpose of the Study:

• To investigate a new family of transformations based on Deep Neural Networks for image watermarking.
• To assess the feasibility and performance of deep learning-based transformations for zero-bit watermarking.
• To evaluate the robustness and quality of watermarked images using advanced computer vision techniques.

Main Methods:

• Feature vector extraction using Deep Neural Networks trained for classification tasks.
• Embedding watermarks by modifying extracted deep features.
• Utilizing adversarial sample literature for inverse transformations to project features back into image space.
• Employing aggregation schemes with weak geometry and retraining with augmented datasets.

Main Results:

• Demonstrated feasibility of Deep Neural Network-based transformations for zero-bit image watermarking.
• Achieved good quality of watermarked images with intrinsic robustness.
• Showcased improved resilience against classical image processing attacks through advanced computer vision methods.

Conclusions:

• Deep Neural Network-based transformations offer a promising alternative to traditional methods in image watermarking.
• The proposed approach provides inherent robustness and maintains high watermarked image quality.
• Further advancements in computer vision can enhance the security and effectiveness of digital watermarking systems.