TL;DR
This paper introduces a novel Reversible Data Hiding in Encrypted Images (RDHEI) method that leverages a Median Edge Detector (MED) for pixel prediction and two's complement encoding for prediction errors. This approach achieves higher embedding rates and improved security by effectively reserving room in encrypted images without compromising the original image's integrity.
What Problem Does This Paper Address?
In the era of cloud storage, users increasingly encrypt their images to protect privacy before uploading them. However, cloud servers need to manage these encrypted images, often by embedding auxiliary data (like authentication or annotation) for efficient retrieval and verification. This creates a challenge: how to embed data into encrypted images in a way that allows both the secret data and the original image to be perfectly recovered without any distortion. Traditional data hiding methods can cause permanent image distortion, which is unacceptable in critical fields like medicine or military applications. Existing RDHEI methods have limitations, such as low embedding rates or the risk of leaking original image information to the data hider through shared auxiliary data. This paper aims to address these issues by proposing a more efficient and secure RDHEI method.
Methods
The proposed method operates in three phases:
- Content Owner Phase:
- Two's Complement Generation and Labeling: The Median Edge Detector (MED) is used to predict pixel values. The difference between the original pixel and its predicted value (prediction error) is calculated. These prediction errors are then encoded using an adaptive-length two's complement. Pixels whose prediction errors fall within a defined interval (determined by parameter 'a') are classified as "labeled pixels."
- Label Map Generation: A binary label map is created to distinguish between labeled and unlabeled pixels.
- Auxiliary Information Embedding: The label map is compressed using extended run-length coding. To embed this compressed map and other necessary parameters (like 'a' and coordinates), an MSB rearrangement method is employed. This involves extracting the (8-a)-bit MSBs of unlabeled pixels and combining them with the compressed label map. This combined bitstream is then embedded into the (8-a)-bit MSBs of pixels in the image, creating a "processed image."
- Image Encryption: The processed image is encrypted using a chaotic encryption system with an image encryption key ($K_e$). Crucially, the bits used for parameters are not encrypted to allow the data hider access.
- Data Hider Phase:
- Secret Data Embedding: The data hider receives the encrypted image. Using a data hiding key ($K_d$), they extract the parameters 'a' and coordinates from the unencrypted reference pixels. The reserved room (identified by these parameters) is then used to embed the encrypted secret data.
- Receiver Phase:
- Data Extraction: If the receiver has the data hiding key ($K_d$), they can extract the secret data from the marked encrypted image.
- Image Recovery: If the receiver has the image encryption key ($K_e$), they first decrypt the image to obtain the processed image. They then extract the parameters 'a', coordinates, and the label map length ($L_m$) from the reference pixels. Using these parameters, they reconstruct the original image losslessly.
Datasets
The paper evaluates its method using six standard test images: Lena, F16, Baboon, Tiffany, Airplane, and Man. For broader validation and comparison with existing methods, it also uses three public datasets:
- UCID (Uncompressed Color Image Database)
- BOSSbase (Break Our Steganographic System)
- BOWS-2
Key Insights and Ideas
- MED for Efficient Prediction: Using the Median Edge Detector (MED) for pixel prediction is a smart way to exploit spatial correlation in natural images, leading to prediction errors that are more concentrated and thus easier to encode efficiently.
- Adaptive Two's Complement: Employing an adaptive-length two's complement for encoding prediction errors allows for flexible representation of different error ranges, maximizing the use of available space.
- Label Map for Distinction: The label map is a critical component that enables the distinction between pixels that can be modified (labeled) and those that must remain untouched (unlabeled), ensuring reversibility.
- MSB Rearrangement for Security: Embedding the label map and parameters in the MSBs of pixels, combined with encryption, prevents the data hider from gaining access to the original image's spatial information, thus enhancing security.
- Parameter Balancing for ER: The choice of the parameter 'a' (controlling the two's complement interval) is crucial. There's a trade-off: larger 'a' means more labeled pixels but smaller reserved room per pixel. The paper demonstrates that an optimal 'a' (around 4) balances these factors to achieve maximum embedding rate (ER).
Key Results
- The proposed method achieves higher embedding rates (ER) compared to previous state-of-the-art RDHEI methods. For example, on the Lena image with $a=4$, the ER reaches 2.864 bpp.
- On the UCID, BOSSbase, and BOWS-2 datasets, the average ERs are 2.58 bpp, 3.04 bpp, and 2.94 bpp, respectively, outperforming existing methods.
- The method demonstrates strong security, as evidenced by the uniform distribution of the encrypted image's histogram, making statistical analysis difficult.
- The recovered image is completely reversible, with a PSNR of +∞ dB between the original and recovered Lena images.
- The method shows good performance on both smooth and rough images.
Limitations and Future Work
- Parameter 'a' Optimization: The paper notes that the optimal value of 'a' varies for different images. While $a=4$ is suggested as a good general choice, further research could explore more adaptive methods for selecting 'a' based on image characteristics.
- Improving ER: The authors suggest two avenues for future work to further increase the ER:
- Developing more efficient coding schemes for labeling prediction errors.
- Improving the compression method for auxiliary information (like the label map).
- Auxiliary Information Reduction: While the MSB rearrangement method enhances security, reducing the amount of auxiliary information needed could further boost embedding capacity.
Summary updated