Scholarlink

<p>The rapid increase in digital content distribution calls for effective watermarking techniques to ensure protection against copyright infringement, imperceptibility, and resilience to adversarial attacks. Current spatial and frequency–domain techniques are not robust against compression, geometric, and hostile attacks, reflecting a deep chasm in secure watermarking techniques. To counter this, a new hybrid watermarking platform that integrates deep learning, frequency-domain transformations, and metaheuristic optimization algorithms has been proposed to improve watermark robustness and security. This method employs a Fractional Discrete Wavelet Transform (FrDWT) to enable multiresolution frequency decomposition, a Dual Attention Vision Transformer (DA-ViT) for precise identification of optimal embedding points, and an African Vultures Optimization Algorithm (AVOA) to enable adaptive power embedding. This approach employs Chaos-Based Watermark Encryption (CBWE) for security enhancement, and the watermark extraction process is performed using Neural Diffusion Model-Based Watermark Extraction (NDM-WE) with the Satin Bowerbird Optimizer (SBO) for high-fidelity restoration. Benchmark test images, namely, Airplane, Baboon, Boats, Goldhill, Lena, Peppers, Sailboat, and Zelda are used for a comprehensive evaluation. The proposed method demonstrates an average PSNR improvement of 15.1%, SSIM enhancement of 1.2%, and Normalized Correlation (NC) gain of 1.6% compared to the existing techniques, such as CNN-Watermarking, Transformer-based methods, and Hybrid DWT-CNN approaches. Additionally, the system achieves the fastest convergence with the minimum fitness value of 0.0123, which is 67.5% better than that of traditional optimization methods, and it exhibits superior computational efficiency, with an average processing time of 1.89 s (up to 59.3% faster).The results confirm the superiority of the proposed method in watermark embedding and extraction optimization by maintaining high imperceptibility, strong attack resistance, and improved security.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DA-ViT: Deep Learning and Frequency–Domain Hybrid Watermarking with Attention-Based Transformers and Diffusion Models

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Rupavathi Natarajan,
  • Kavitha Manickam

摘要

The rapid increase in digital content distribution calls for effective watermarking techniques to ensure protection against copyright infringement, imperceptibility, and resilience to adversarial attacks. Current spatial and frequency–domain techniques are not robust against compression, geometric, and hostile attacks, reflecting a deep chasm in secure watermarking techniques. To counter this, a new hybrid watermarking platform that integrates deep learning, frequency-domain transformations, and metaheuristic optimization algorithms has been proposed to improve watermark robustness and security. This method employs a Fractional Discrete Wavelet Transform (FrDWT) to enable multiresolution frequency decomposition, a Dual Attention Vision Transformer (DA-ViT) for precise identification of optimal embedding points, and an African Vultures Optimization Algorithm (AVOA) to enable adaptive power embedding. This approach employs Chaos-Based Watermark Encryption (CBWE) for security enhancement, and the watermark extraction process is performed using Neural Diffusion Model-Based Watermark Extraction (NDM-WE) with the Satin Bowerbird Optimizer (SBO) for high-fidelity restoration. Benchmark test images, namely, Airplane, Baboon, Boats, Goldhill, Lena, Peppers, Sailboat, and Zelda are used for a comprehensive evaluation. The proposed method demonstrates an average PSNR improvement of 15.1%, SSIM enhancement of 1.2%, and Normalized Correlation (NC) gain of 1.6% compared to the existing techniques, such as CNN-Watermarking, Transformer-based methods, and Hybrid DWT-CNN approaches. Additionally, the system achieves the fastest convergence with the minimum fitness value of 0.0123, which is 67.5% better than that of traditional optimization methods, and it exhibits superior computational efficiency, with an average processing time of 1.89 s (up to 59.3% faster).The results confirm the superiority of the proposed method in watermark embedding and extraction optimization by maintaining high imperceptibility, strong attack resistance, and improved security.