Maximizing Strength of Digital Watermarks using Fuzzy Logic

In this paper, we propose a novel digital watermarking scheme in DCT domain based fuzzy inference system and the human visual system to adapt the embedding strength of different blocks. Firstly, the original image is divided into some 8 \times 8 blocks, and then fuzzy inference system according to different textural features and luminance of each block decide adaptively different embedding strengths. The watermark detection adopts correlation technology. Experimental results show that the proposed scheme has good imperceptibility and high robustness to common image processing operators.

💡 Research Summary

The paper proposes an adaptive digital watermarking scheme that leverages a fuzzy inference system (FIS) together with a Human Visual System (HVS) model to determine the embedding strength of each 8×8 DCT block in an image. The authors first divide the host image into 8×8 blocks and extract five texture features—Angular Second Moment, Contrast, Correlation, Variance, and Entropy—from a gray‑level co‑occurrence matrix for each block. Along with the block’s average luminance, these features serve as inputs to a Mamdani‑type fuzzy system. Each input is fuzzified into three linguistic terms (Low, Average, High) using membership functions generated with MATLAB’s ANFIS toolbox, and a rule base is automatically derived. The fuzzy output is a weighting factor α that controls how two selected middle‑frequency DCT coefficients within the block are modified: one coefficient is increased by α while the other is decreased, or vice‑versa. This “opposite‑sign” embedding maintains local energy balance and exploits the fact that the HVS is less sensitive to changes in textured or high‑luminance regions, allowing stronger embedding where visual masking is higher.

Detection is blind; the original image is not required. The possibly attacked image undergoes the same DCT transform, the same pair of coefficients is examined, and a correlation operation determines whether the embedded bit is a ‘0’ or ‘1’. Two secret keys are used: one to specify the positions of the coefficient pair (selected from quantization‑identical locations) and another to indicate which blocks contain the watermark.

To evaluate visual quality, the authors introduce a weighted Peak Signal‑to‑Noise Ratio (wPSNR) that incorporates a weighted MSE (wMSE) reflecting HVS sensitivity: errors in homogeneous regions are penalized more heavily than those in heterogeneous regions. They compare both PSNR and wPSNR across a set of 800 medical images (256×256 pixels) taken from a DICOM library. Robustness is tested primarily against JPEG compression with quality factors ranging from 90 % down to 10 %. Results show that wPSNR consistently exceeds PSNR, indicating better perceptual fidelity, while the correlation‑based detector successfully recovers the watermark even at low quality factors (e.g., 35 %). The paper reports mean PSNR values around 40 dB and demonstrates that the fuzzy‑controlled embedding yields higher robustness than a non‑adaptive baseline.

The contribution lies in integrating texture‑based HVS masking with fuzzy reasoning to adapt embedding strength on a per‑block basis, moving beyond simple luminance‑only masking. However, the study has several limitations. The fuzzy rule set is generated from the same dataset used for testing, raising concerns about over‑fitting and generalization to other image domains. The experimental evaluation focuses mainly on JPEG compression; other common attacks such as Gaussian noise, median filtering, scaling, rotation, or geometric cropping are not examined. Comparative analysis with recent state‑of‑the‑art watermarking methods, especially those employing deep learning or more sophisticated perceptual models, is absent, making it difficult to gauge the relative advantage of the proposed approach. Security considerations are also limited to the description of two keys; the paper does not discuss key distribution, resistance to key‑guessing attacks, or the impact of embedding strength on payload capacity. Finally, the manuscript contains numerous typographical and formatting errors, which could hinder reproducibility without access to the authors’ source code and detailed parameter settings.

In summary, the paper presents a conceptually sound method that uses fuzzy logic to tailor watermark strength according to local texture and luminance, achieving a favorable trade‑off between imperceptibility and robustness for medical images. While the experimental results are promising, further work is needed to validate the approach under a broader set of attacks, to compare it against contemporary techniques, and to provide a more rigorous security analysis before the method can be considered ready for practical deployment in sensitive medical imaging environments.