The Berbomthum remote access trojan was reported by TrendMicro to accept commands from a Twitter account created in 2017. The account used harmless malicious memes that used steganography as a technique for communication between the malware and programmers or attackers6. The calibration technique is also used to understand the approximation of the cover image from the stego image. The research uses Support Vector Machine (SVM) and its evolutionary version Support Vector Machine-Particle Swarm Optimization (SVM-PSO) to evaluate the classification result under the same conditions.
This would improve the accuracy of the analysis as the image is analyzed in terms of blocks rather than a single frame. It is because of the wide real-time use of the format in the Internet transmission due to its high compression ratio. The statistical features include: the first-order features, the second-order features, and the Markovian features.
In this paper, we discuss and evaluate the performance of steganographic algorithms from the spatial and transformation domains. Let P(X) be the probability density, where X is the grayscale of the image and Gn is the number of pixels with a grayscale value. Then, the probability density of the gray-level image X is given by Eq.
In the spatial domain, the secret message is embedded directly into the image pixel value of the cover image. The message to be hidden is embedded by replacing the LSB of the image byte sequentially or using a pseudo-random key1,54. If the length of the message is less than the cover pixel length of the cover image, the pseudorandom permutation ensures that the changes are spread evenly across the image.
The steganographic algorithm that makes use of the DCT and is used in this research is F5. This histogram gives an idea of the total number of times a particular coefficient will appear in the entire image in all the blocks in the transformation. Figures 8 and 9 below give an overview of the dual histogram of a sample stego image and its equivalent in a calibrated stego image.
Once both images are available, the image characteristics can be compared to the L1 norm.
Results
Do t Radiale Polinoom Multikwadri c Epanechnikov ANOVA Do t Radiale Polinoom Multiquadri c Epanechnikov ANOVA Do t Radiale Polinoom Multiquadri c Epanechnikov ANOVA Do t Radiale Polinoom Multiquadri c Epanechnikov ANOV A.
Classification of uncalibrated images using SVM
Do t Radial Polynomial Multiquadri c Epanechnikov ANOV A Do t Radial Polynomial Multiquadri c Epanechnikov ANOV A Do t Radial Polynomial Multiquadri c Epanechnikov ANOV A Do t Radial Polynomial Multiquadri c Epanechnikov ANOV A.
Classification of calibrated images using SVM
For LSB replacement, ANOVA and radial kernels in linear sampling in Table 5 have a percentage ranging from 33 to 36. For LSB matching, the pattern is followed with a better classification given by the polynomial of its mixed sampling. The PVD scheme has a very good classification, with ANOVA giving a classification of 99.8 in its linear sampling.
For the F5, the classification percentage of point and multiquadries is from 29 to 38% for shuffled, stratified and automatic sampling. The good classification rate in this scheme is exhibited by the multiquadric kernel in linear sampling. In LSB substitution in Table 7, the point and multiquadric kernels give a classification for all samplings in the range of 31% to 39%.
The PVD scheme gives Epanechnikov's classification and radian in the range of 25 to 29 for all sampling.
Classification of uncalibrated images using SVM-PSO
Do t Radiale Polinoom Multikwadri c Epanechnikov ANOVA Do t Radiale Polinoom Multiquadri c Epanechnikov ANOVA Do t Radiale Polinoom Multikwadri c Epanechnikov ANOV A Do t Radiale Polinoom Multiquadri c Epanechnikov ANOVA.
Classification of calibrated images using SVM-PSO
Do t Polinom radial Multiquadri c Epanechnikov ANOV A Do t polinom radial Multiquadri c Epanechnikov ANOVA Do t Polinom radial Multiquadri c Epanechnikov ANOV A Do t Polinom radial Multiquadri c Epanechnikov ANOV A.
Classification of uncalibrated images with cross-validation using SVM
In LSB replacement and LSB matching in Table 8, most kernels with different sampling have classification percentage between 75 and 80%. LSB replacement and LSB matching have the most classification for all cores and sampling to be between 75 and 80%. The best classification of LSB replacement and LSB matching is given by multiquadric in linear sampling.
For the F5 steg anographic scheme, almost all kernels and sampling give a good classification rate, but a better classification is given by radial, multiquadric and Epanechnikov kernels in linear sampling. In LSB replacement and LSB matching in Table 9, the point kernel gives an accuracy between 38 and 39%. The F5 algorithm follows the same pattern as LSB replacement and LSB matching, with Epanechnikov giving good classification results.
Classification of calibrated images with cross-validation using SVM
Up to t Radial Polynomial Multiquadric Epanečnik ANOVA Up to t Radial Polynomial Multiquadric Epanečnik ANOVA Up to t Radial Polynomial Multiquadric Epanečnik ANOVA Up to t Radial Polynomial Multiquadric Epanečnik ANOVA.
Classification of uncalibrated images with cross-validation using SVM-PSO
The better classification is given by the polynomial kernel in its stratified and automatic sampling.
Conclusion
Future scope
Classification of calibrated images with cross-validation using SVM-PSO
Data availability
A Prediction of Precipitation Data Based on Support Vector Machine and Particle Swarm Optimization (PSO‑SVM) Algorithms (MDPI, 2017). Feature-based stegan analysis for JPEG images and its implications for future steganographic scheme design. Steganography and Steganalysis of JPEG Images: A Statistical Approach to Information Hiding and Detection (University of Florida, 2011).
An efficient and robust bat algorithm combining adversarial learning and whale optimization algorithm. A hybrid PSO optimized SVM-based model for predicting successful growth style of Spirulina platensis from raceway experiment data. Effect of features selected by principal component analysis in feature-based stegaanalysis on calibrated and uncalibrated images.
Landslide susceptibility mapping based on Support Vector Machine: A case study on natural slopes in Hong Kong, China.
Author contributions
Competing interests
Additional information
