TL;DR
This paper introduces IEWNet, a novel multi-scale robust watermarking network designed to withstand Infrared Image Enhancement (IRE) attacks. By employing a multi-scale architecture and incorporating specialized noise and enhancement modules, IEWNet significantly improves watermark extraction accuracy and imperceptibility in IR images subjected to various enhancement techniques.
What Problem Does This Paper Address?
Infrared (IR) images are crucial for applications ranging from medical diagnostics to disaster detection. However, IR images often suffer from blurred edges and limited detail due to thermal diffusion. To address this, Infrared Enhancement (IRE) algorithms are commonly applied. While IRE improves image quality, it often corrupts embedded watermarks, hindering copyright protection and data integrity. Existing watermarking methods struggle to maintain robustness against these IRE attacks, creating a need for specialized solutions.
Methods
The proposed IEWNet utilizes an auto-encoder architecture with several key innovations:
- Multi-Scale Feature Extraction: IEWNet employs a multi-scale neural network structure to embed watermarks at different resolutions of the IR image. This allows for more comprehensive feature extraction and selection of optimal embedding locations.
- Undecimated Dual Tree Complex Wavelet Transform (UDTCWT): A UDTCWT network is used for preprocessing, extracting features from the IR image that are then fed into the encoder.
- Noise and Enhancement Modules: A crucial component is the "noise network" placed between the encoder and decoder. This network simulates IRE attacks by incorporating:
- A noise layer trained to mimic eight traditional IRE algorithms (e.g., HE, mean filter, median filter).
-
An enhancement sub-network trained to mimic four machine learning-based IRE algorithms. These modules allow the network to learn and adapt to various IRE attack patterns, thereby improving robustness.
-
Encoder-Decoder Architecture: The core of IEWNet is an encoder that embeds the watermark and a decoder that extracts it. Both utilize a multi-scale structure with attention mechanisms to enhance feature extraction.
Datasets
The paper uses six public IR image datasets for training and evaluation:
- GTOT
- Road-Scene
- TNO
- LasHeR
- RGBT210
- RGBT234
These datasets contain a variety of IR images captured under different conditions, including indoor and outdoor environments, day and night. The images were preprocessed by random filtering, cropping, or scaling to a size of 128x128 pixels.
Key Insights and Ideas
- IRE as a Primary Attack Vector: The paper correctly identifies IRE as a significant threat to IR image watermarking, motivating the development of specialized defenses.
- Simulating Attacks for Robustness: Instead of directly defending against every possible IRE algorithm, IEWNet learns to simulate these attacks through dedicated noise and enhancement modules. This allows the network to generalize better to unseen IRE variations.
- Multi-Scale is Key for IR: The inherent characteristics of IR images (blurred edges, thermal diffusion) make multi-scale processing particularly effective for both feature extraction and watermark embedding.
- Trade-off between Robustness and Imperceptibility: While not explicitly detailed as a limitation, the paper's focus on robustness against IRE attacks implies a careful balance is needed to maintain imperceptibility, which is evaluated using PSNR and SSIM.
Key Results
- Imperceptibility: The proposed IEWNet achieves high imperceptibility, with Peak Signal to Noise Ratio (PSNR) values generally exceeding 40 dB across six test datasets. Structural Similarity Index Measure (SSIM) values also remained high, indicating minimal visual degradation.
- Robustness: IEWNet demonstrates superior robustness against both traditional and machine learning-based IRE attacks compared to existing state-of-the-art watermarking algorithms.
- Average Bit Error Ratio (BER) values were significantly lower than comparative methods. For instance, on the GTOT dataset, IEWNet achieved an average BER of 1.50%, outperforming others which ranged from 1.81% to 2.64%.
- The noise network component was shown to significantly improve robustness against IRE attacks.
Limitations and Future Work
- IRE Algorithm Identification: The paper acknowledges that determining the specific IRE algorithm applied to an enhanced IR image is a challenge. Future work should focus on developing algorithms that can quickly and accurately analyze the type of IRE applied.
- Broader Attack Scenarios: While IEWNet is robust against IRE attacks, the authors suggest future research should consider more complex industrial scenarios, including special noise generated during transmission and combined attacks from multiple IR processing algorithms.
- Algorithm Categorization: A more precise categorization of IRE algorithms is recommended for developing more targeted and effective watermarking solutions.
Summary updated