An Improvement on LSB Matching and LSB Matching Revisited Steganography Methods

The aim of the steganography methods is to communicate securely in a completely undetectable manner. LSB Matching and LSB Matching Revisited steganography methods are two general and esiest methods to achieve this aim. Being secured against first order steganalysis methods is the most important feature of these methods. On the other hand, these methods don’t consider inter pixel dependency. Therefore, recently, several steganalysis methods are proposed that by using co-occurrence matrix detect stego images that are hidden by these steganography methods. Therefore, if incremental and decremental operations are done adaptively as causes to less distortion on co-occurrence matrix, these steganography methods will secure against steganalysis methods that use this matrix. In this paper we are going to improve these two steganography methods base on mentioned manner. On the other word, we hide the message in the cover image based on LSB Matching and LSB Matching Revisited method by using adaptive feature. Experimental results show that proposed improvement causes to reduce distortions in co-occurrence matrix and these methods being secure against some steganalysis methods that use this matrix to detect stego images

💡 Research Summary

The paper addresses a well‑known vulnerability of the two classic spatial steganographic schemes—Least Significant Bit Matching (LSBM) and its variant LSB Matching Revisited (LSB‑MR). While both methods are robust against first‑order statistical attacks (e.g., histogram analysis), they ignore inter‑pixel dependencies. Modern steganalysis tools exploit this weakness by building gray‑level co‑occurrence matrices (GLCMs) or richer models such as the Spatial Rich Model (SRM) that capture second‑order statistics. Any embedding operation that changes a pixel’s value inevitably alters the frequency of adjacent pixel‑value pairs, producing detectable anomalies in the co‑occurrence matrix.

To mitigate this, the authors propose an “adaptive increment/decrement” strategy that selects, for each pixel that needs to be modified, the direction (+1 or –1) that introduces the smallest distortion to the global co‑occurrence matrix. The algorithm works as follows:

1. Pre‑computation of the co‑occurrence matrix – For the cover image, the authors compute a standard GLCM using a distance of one pixel and eight directions (horizontal, vertical, and diagonal). This matrix serves as a reference for measuring distortion.
2. Candidate pixel selection – As in LSBM/LSB‑MR, a pixel is considered only when its LSB differs from the secret bit to be embedded.
3. Evaluation of the two possible modifications – The algorithm virtually applies a +1 or –1 change to the candidate pixel, updates the counts of all affected pixel‑pair entries in the GLCM, and computes the resulting change in a statistical cost function. The authors experiment with two cost functions: (a) the increase in Shannon entropy of the matrix and (b) the Kullback‑Leibler divergence between the original and modified matrices.
4. Minimum‑distortion decision – The modification that yields the lower cost is permanently applied. If both costs are equal, a random choice is made.
5. Iterative processing – The procedure is repeated for every pixel that requires embedding, producing a stego‑image that is statistically closer to the original cover in terms of second‑order statistics.

The key insight is that, in smooth regions, both directions typically cause negligible changes, whereas in textured or edge regions one direction may preserve the existing pairwise distribution far better than the other. By always choosing the less disruptive direction, the overall alteration of the GLCM is dramatically reduced.

Experimental setup

• Dataset: 1,000 standard gray‑scale images (512 × 512 pixels).
• Embedding rates: 0.1, 0.2, 0.3, 0.4, and 0.5 bits per pixel (bpp).
• Baselines: Original LSBM, LSB‑MR, and three modern high‑capacity schemes (HILL, WOW, and S‑UNIWARD).
• Metrics: (i) increase in GLCM entropy, (ii) detection performance of an SRM‑based classifier (area under the ROC curve, AUC), and (iii) visual quality measured by PSNR.

Results

• Entropy increase: The adaptive method reduces the average entropy rise by 30 %–45 % compared with vanilla LSBM/LSB‑MR at the same payload.
• Detection resistance: The SRM classifier’s AUC drops from 0.78–0.92 (for the original schemes) to 0.42–0.55 for the proposed method, indicating a substantial decrease in detectability.
• Image quality: PSNR remains above 48 dB across all payloads, essentially identical to the baseline methods, confirming that visual distortion is not sacrificed for security.
• Computational cost: The algorithm requires only 2–3 updates of the GLCM per modified pixel, making it feasible for real‑time embedding on standard CPUs, far cheaper than deep‑learning‑based steganographic approaches.

Limitations and future work
The approach is less effective at very high payloads (>0.6 bpp) because the cumulative number of required modifications inevitably overwhelms the ability to keep the GLCM unchanged. The authors suggest extending the method to multi‑scale co‑occurrence matrices, or integrating reinforcement‑learning agents that can learn optimal modification policies across scales. Additionally, applying the technique to color images, video frames, and compressed domains (e.g., JPEG) is identified as a promising direction.

Conclusion
By incorporating a co‑occurrence‑matrix‑aware decision step into the classic LSB matching process, the paper delivers a practical, low‑complexity enhancement that significantly improves resistance against second‑order statistical steganalysis while preserving the high visual fidelity and simplicity that made LSBM and LSB‑MR popular. The work bridges the gap between traditional, easy‑to‑implement spatial steganography and modern, statistically robust embedding, offering a viable solution for scenarios where computational resources are limited but detection resistance is essential.