سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / سومین کنفرانس بین المللی مهندسی و علوم کامپیوتر و اولین کنفرانس ملی سامانه های هوشمند، کلان داده و رایانش پیشرفته

A Curriculum Learning Approach for Robust Blind Deep Image Watermarking against Salt and Pepper Noise

نویسندگان :

Mohammad Ebadi¹ Alireza Naghsh² Seyed Mahmoud Daneshvar³

1- دانشگاه آزاد اسلامی واحد نجف آباد 2- دانشگاه آزاد اسلامی واحد نجف آباد 3- دانشگاه آزاد اسلامی واحد نجف آباد

کلمات کلیدی :

Robust Watermarking،Salt and Pepper Noise،Convolutional Auto-Encoder،Curriculum Learning

چکیده :

The escalating need for secure and covert communication channels necessitates the development of robust watermarking techniques that can withstand significant channel noise. Conventional methods often exhibit fragility when subjected to non-Gaussian noise, such as high-density Salt and Pepper noise. This paper introduces a novel deep learning framework for watermarking, engineered for exceptional resilience against such adversarial conditions. The architecture comprises an end-to-end autoencoder structure built upon convolutional layers, with an encoder to embed a binary message into a grayscale host image and a decoder to extract it. Our primary contribution lies in a novel three-stage curriculum learning strategy that dynamically adjusts loss function weights to optimize for competing objectives sequentially. The training begins in stage 1 by prioritizing message recovery accuracy, subsequently balancing it with image perceptual quality in stage 2, and concludes in stage 3 by fine-tuning for imperceptibility. Robustness is explicitly learned by introducing randomly sampled noise as a disruptive layer between the encoder and decoder during the training loop. Extensive evaluations on the MNIST dataset demonstrate the framework's superior performance. The proposed model achieves a near-zero Bit Error Rate for noise densities up to 90% and maintains successful message extraction. Concurrently, the system preserves high perceptual fidelity, with an average Peak Signal-to-Noise Ratio exceeding 44 dB, Structural Similarity Index Measure above 0.90. This work confirms that a structured curriculum learning approach, combined with in-loop adversarial training, is a highly effective methodology for creating watermarking systems with state-of-the-art robustness, enabling reliable covert communication over volatile channels.