A Feature Selection Method for Steganalysis Based on Clonal Selection Algorithm

Abstract

The present invention discloses a steganalysis feature selection method based on the clonal selection algorithm. Aiming at the problems that the dimension of the classical rich model steganalysis features is too high and contains a large number of redundant features, the present invention realizes the reduction of the high-dimensional steganalysis feature set based on the clonal selection algorithm. By designing an affinity function Af of the features to calculate the affinity value aff of the features, and sorting the features according to the level of the affinity value. Through T rounds of iteration, the best feature subset F is screened out from the features in a specific niche, and through mutation operation, the mutated feature subset Fm is obtained; the classification performances of the feature subset F and the mutated feature subset Fm are compared to obtain the final feature subset C of this niche; after T rounds of iteration, the target feature subset C1 is obtained. The method of the present invention is easy to calculate and implement, and can greatly reduce the feature dimension while ensuring the detection accuracy, and is effective for steganalysis features in both the spatial domain and the JPEG domain, and has practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of information hiding, and in particular to a steganalysis feature selection method based on a clonal selection algorithm. Background Art

[0002] Image steganography embeds secret information into the redundant parts of an image, and there is no obvious visual difference before and after embedding the secret information, so as to achieve the purpose of covert communication. Steganalysis aims to detect the steganographic signal in an image to determine whether the image contains secret information. Adaptive steganography algorithms have greatly improved the security of steganography, but also brought great challenges to steganalysis. Therefore, researchers introduce complex high-dimensional features into steganalysis to improve the detection accuracy. In the spatial domain, the literature (J. Fridrich and J. Kodovsky, "Rich models for steganalysis of digital images," IEEE Transactions on Information Forensics and Security, vol. 7, no. 3, pp. 868-882, 2012) proposed the Spatial Rich Model (SRM). The SRM feature has a total of 34,671 dimensions and has high detection performance, which is widely used in current image steganalysis. In the JPEG domain, the literature (J. Kodovsky and J. Fridrich, "Steganalysis of JPEG images using rich models," Proceedings of SPIE - The International Society for Optical Engineering, vol. 8303, 8303A, 2012) borrowed the co-occurrence matrix idea of SRM and proposed the JPEG Rich Model (JRM) feature of 11,255 dimensions.

[0003] Although high-dimensional steganalysis features show better performance in detecting steganographic signals, they consume a large amount of time and computational cost in feature extraction and classifier training. In addition, there are a large number of redundant features in high-dimensional features. Removing redundant features will not reduce the detection accuracy, and may even slightly improve the detection accuracy.

[0004] In recent years, with the rapid development of deep learning, steganalysis methods based on deep learning have been widely studied, and quite a number of deep steganalysis networks have been proposed. However, most deep steganalysis networks, such as SRNet proposed in the literature (M. Boroumand, M. Chen, and J. Fridrich, “Deep residual network for steganalysis of digital images,” IEEE Transactions on Information Forensics and Security, vol. 14, no. 5, pp. 1181–1193, 2019) and Zhu-Net proposed in the literature (R. Zhang, F. Zhu, J. Liu, and G. Liu, “Depth-wise separable convolutions and multi-level pooling for an efficient spatial CNN-based steganalysis,” IEEE Transactions on Information Forensics and Security, vol. 15, pp. 1138–1150, 2020), focus on constructing deeper and more complex networks to improve detection performance, which results in overly long model training time and huge computational costs. In addition, deep steganalysis networks also suffer from the problem of insufficient generalization ability, leading to a lack of generality when facing complex and variable images in the real world.

[0005] With the continuous development of computational intelligence, the branch of the Clonal Selection Algorithm (CSA) has attracted much attention from research scholars. CSA uses a memory mechanism to ensure rapid convergence to the global optimal solution, has excellent search capabilities, and has achieved good performance in fields such as pattern recognition and classification problems. In this work, the affinity value of features is calculated by designing an affinity function for features, the features are sorted according to the affinity value of the features, and the number of iterations is calculated. By calculating the number of similar features for the features in each iteration, the screening of the features is completed. Finally, the feature subset obtained by screening is mutated, and the classification effect is compared. The target feature subset C1 is determined according to the classification effect. Aiming at the problems of too high dimensionality and a large number of redundant features in high-dimensional steganalysis features, the present invention performs feature reduction on the high-dimensional steganalysis feature set, and uses the affinity mechanism and mutation mechanism in CSA to complete the feature selection operation, greatly reducing the feature dimensionality while ensuring the detection accuracy. Summary of the Invention

[0006] In order to reduce the feature dimension by screening out redundant features in the high-dimensional feature set while keeping the detection accuracy unaffected, the present invention proposes a steganography analysis feature selection based on the clonal selection algorithm. This method can calculate the affinity value of features using an affinity function, obtain a feature subset F using a feature screening algorithm, and at the same time perform a mutation operation on the feature subset using the mutation operation in the clonal selection algorithm to obtain a mutated feature subset Fm. By comparing the classification effects of the feature subset F and the mutated feature subset Fm, the target feature subset C1 is determined. This method can greatly reduce the feature dimension while maintaining the detection accuracy, reducing time and computational costs.

[0007] The technical solution of the present invention is implemented as follows:

[0008] A steganography analysis feature selection method based on the clonal selection algorithm, which realizes the reduction of the high-dimensional steganography analysis feature set based on the clonal selection algorithm, reduces the feature dimension by screening out redundant features in the high-dimensional feature set, designs an affinity function Af of features to calculate the affinity value aff of features, sorts the features according to the affinity value, and selects the best feature subset F from the features in a specific niche through T rounds of iteration, and performs a mutation operation on the feature subset F, that is, randomly selects a certain number of similar features to be mutated and adds them to the feature subset F to obtain a mutated feature subset Fm; compares the classification performance of the feature subset F and the mutated feature subset Fm to obtain the final feature subset C of this niche; obtains the target feature subset C1 after T rounds of iteration; includes the following main steps:

[0009] Step 1, use the feature extraction function E to extract the steganography analysis features of the carrier image dataset X and the stego-image dataset Y respectively as the feature subset Fc to be reduced and the feature subset Fs to be reduced, and calculate the affinity value of each dimension of feature in the feature subset to be reduced; first divide the feature subset to be reduced into n {n ∈ N +}, calculate the affinity value of the features in each part; perform maximum-minimum normalization on each dimension of feature, and calculate the average value aver fc and aver fs and the standard deviation std fc and std fs in the corresponding carrier image or stego-image class respectively, and finally calculate the affinity value aff of the feature according to the calculated average value and standard deviation:

[0010]

[0011] In the formula, aff is the affinity value of each dimension of feature in each part of the feature set, and finally obtain the feature affinity value set of each feature subset to be reduced;

[0012] Step 2: According to the feature affinity value aff obtained in Step 1, sort each feature subset to be reduced according to the value, classify the sorted features into niches, and determine the features and the number of iterations T in each niche; the number of iterations T is dynamically determined according to the feature affinity value aff:

[0013]

[0014] where aff b and aff l represent the maximum and minimum values of the feature affinity of each in n {n ∈ N +}, respectively, and niche refers to the iteration radius;

[0015] Step 3: Perform feature selection on the features in each iteration process; first calculate the similarity between features, and the similarity between feature k and feature g is:

[0016]

[0017] For the calculated dist kg , find all features less than the similarity threshold S t {0 < S t < 1} as the similar features of feature k, and count the number of similar features of each feature; select the feature with the fewest similar features, and use its similar features as redundant features to form a redundant similar feature set SF until the redundant similar feature set SF is the same as all features in a specific niche; finally, output the feature subset F;

[0018] Step 4: Mutate the feature subset F obtained in Step 3, randomly select a certain number of mutated features from the redundant similar feature set SF and add them to the feature subset F to obtain the mutated feature subset Fm; the number of mutated features is:

[0019] mut num = min{P m * aver F , len(F s )}

[0020] where P m is the preset number of mutated features, which can be dynamically adjusted according to the feature dimension or affinity value, aver F is the average value of the affinity values of all features in the feature subset F; len(F s ) is the number of features in the redundant similar feature set SF of F; set the selected number of mutated features mut num to P m * aver F and len(F s) in the minimum value;

[0021] Step 5: Based on the feature subset F obtained in step 3 and the variant feature subset Fm obtained in step 4, the detection accuracy of F and Fm in the ensemble classifier M is compared, and the larger value is taken as the final feature subset C of this iteration;

[0022] Step 6: Based on the number of iterations T obtained in step 2, if the iteration is still in progress, continue to screen the features and repeat steps 3 to 5. After T rounds of iterations, the target feature subset C1 is obtained.

[0023] Furthermore, the number of iterations T in step 2 can be dynamically adjusted according to the feature dimension and affinity value.

[0024] Furthermore, the similarity dist in step 3 kg According to n{n∈N +} is set dynamically, and the similarity threshold S t {0<S t <1} can also be dynamically adjusted according to specific requirements.

[0025] Furthermore, the number of mutations in step 4 is mut num According to the pre-set P m The affinity value of the feature is adjusted dynamically.

[0026] Furthermore, the method of the present invention can be applied to feature selection of spatial domain steganalysis as well as feature selection of JPEG domain steganalysis.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. This invention utilizes the core concept of the clonal selection algorithm to filter high-dimensional feature sets by introducing niche techniques, affinity functions, and feature mutation operations. This significantly reduces feature dimensionality while maintaining detection accuracy. It is also applicable to steganalysis feature selection in both spatial and JPEG domains.

[0029] 2. The features in the microhabitat are screened through the feature similarity function, which ensures that the classification effect of the screened feature subset is consistent with that of the high-dimensional feature set. The number of mutations is reasonably calculated in the feature mutation operation, which effectively ensures the selection of high-affinity features and the screening of low-affinity features. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the process of the present invention.

[0031] Figure 2 This is a framework diagram of the steganalysis feature selection method based on the clonal selection algorithm of the present invention.

[0032] Figure 3 This is a comparison graph of the experimental results of the SRM features of the present invention on the BOSSbase 1.01 dataset. Specific implementation manners

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0034] Aiming at the problems of too high dimensionality of the high-dimensional feature set and a large number of redundant features in image steganography detection, the present invention proposes a steganalysis feature selection method based on the clonal selection algorithm.

[0035] As Figure 1 and Figure 2 shown, the specific steps of the present invention are as follows:

[0036] First, calculate the affinity value aff of the features through the affinity function Af of the features, sort the features according to the affinity value, and determine the number of iterations T. For the feature set in the niche in the t-th (1 ≤ t ≤ T) round of iteration calculate the similar feature s of each dimension feature k in f , that is, the similarity value dist kg satisfying dist kg < S t of the feature g is denoted as the similar feature s of the feature k f , where S t is the set similarity threshold, and the number of similar features of the feature k is denoted as its number of similar features n. Select the features with the least number of similar features n to form the optimal feature subset F, and the similar features s of all the features in F f form the redundant similar feature set SF of F; perform a mutation operation on the feature subset F, that is, randomly select a certain number, that is, the mutation number mut num , of the similar features s f from the redundant similar feature set SF of F and add them to the feature subset F to obtain the mutated feature subset Fm; compare the classification performance of the feature subset F and the mutated feature subset Fm to obtain the final feature subset C of this niche; repeat the above operations for the features in each round of iteration until the stop condition t > f is satisfied; take the union of the final feature subsets C in each round of iteration as the target feature subset C1 for feature selection; including the following main steps:

[0037] Step 1, use the feature extraction function E to extract the sets Fc and Fs of the feature values of the carrier image dataset X and the stego-image dataset Y, divide the sets Fc and Fs of the feature values into n {n ∈ N +} parts, and for each part of the set of feature values and Calculate the affinity value aff of the features in and for the feature values D of each dimension in cj and D sj Perform maximum - minimum normalization operation to obtain and the set N of the normalized values of all features in cj and N sj and calculate the average value aver cj of D sj and D cj in N sj and N fc and aver fs and the standard deviation std fc and std fs Finally, calculate the affinity value aff of each dimension feature according to the calculated average value and standard deviation:

[0038]

[0039] In the formula, aff is the set of affinity values of each dimension of the feature value and Finally, obtain the set of feature affinity values Aff of the set of each feature value and ; j ;

[0040] Step 2: According to the set of feature affinity values Aff of the set of each feature value and obtained in Step 1, sort the features according to the level of the affinity value aff of each dimension feature; and determine the number of iterations T: j In the formula, aff

[0041]

[0042] represents the maximum and minimum values of the feature affinity of each part in n {n ∈ N b} parts respectively, and niche refers to the iteration radius; after determining the number of iterations T, use T and the feature affinity value aff to determine the feature set in the t - th (1 ≤ t ≤ T) round of iteration niche l The range of the affinity value aff of the features in the feature set +} is: Feature set is:

[0043]

[0044] Step 3. According to the number of iterations T obtained in Step 2, perform feature selection on the feature sets in each round of iteration First, calculate the similarity between features. The similarity value between feature k and feature g is:

[0045]

[0046] According to the set similarity threshold S t {0 < S t < 1}, calculate the similar feature s of feature k f , that is, the feature g that satisfies is recorded as the similar feature s of feature k f , and add feature g to the redundant similar feature set SF of feature k k . The number of features in the redundant similar feature set SF k is recorded as the number of similar features n of feature k. Select the feature k with the fewest number of similar features n to form the optimal feature subset F. Take the union of the redundant similar feature sets SF k of all features in the feature subset F as the redundant similar feature set SF of the feature subset F; until the redundant similar feature set SF contains all the features in; finally, output the feature subset F;

[0047] Step 4. Perform feature mutation on the feature subset F obtained in Step 3, that is, randomly select a certain number, that is, the mutation number mut num , of features from the redundant similar feature set SF of the feature subset F and add them to the feature subset F to obtain the mutated feature subset Fm; the mutation number is:

[0048] mut num = min{P m * aver F , len(F s )}

[0049] In the formula, P m is the preset number of feature mutations, aver F is the average value of the affinity values aff of all features in the feature subset F; len(F s ) is the number of features in the redundant similar feature set SF of F; set the selected number of mutated features mut num to the minimum of P m * aver F and len(F s );

[0050] Step 5: According to the feature subset F obtained in Step 3 and the mutated feature subset Fm obtained in Step 4, compare the classification performances of the feature subset F and the mutated feature subset Fm. If the classification performance of the feature subset F is better than that of Fm, set the feature subset F as the final feature subset C in the t-th round; otherwise, set the mutated feature subset Fm as the final feature subset C in the t-th round.

[0051] Step 6: According to the number of iterations T obtained in Step 2, after one round of iteration, t = t + 1; if still in the iteration process, that is, 1 ≤ t ≤ T is satisfied, continue to screen the features in the t-th round, and repeat Step 3, Step 4, and Step 5 until the stopping condition t > T is satisfied; take the union of the final feature subsets C in each round of iteration as the target feature subset C1 for feature selection.

[0052] The effectiveness of this method is illustrated below with a specific experiment.

[0053] Given a BOSSBase 1.01 dataset, which contains 10,000 grayscale images, set the image size to 256×256 as the carrier image for the experiment. And use the S-UNIWARD steganographic algorithm on the BOSSBase 1.01 dataset to obtain stego-images with 0.2, 0.4, 0.5, 0.8, and 1.0 bpp (bits per pixel) respectively. Use the SRM feature extraction algorithm to extract 34,671-dimensional SRM features from the carrier image and the 5 stego-images. Perform feature selection on the SRM features of the 5 stego-images respectively.

[0054] Step 1: Calculate the normalized eigenvalue, the corresponding average value, and the standard deviation of each dimension of the SRM features, and obtain the affinity value aff of each dimension of the feature in n {n ∈ N +} copies.

[0055] Step 2: Determine the number of iterations T. Before meeting the iteration stop condition, perform feature selection on the features in each round of iteration. First, calculate the similar features of the features in this round of iteration, and select the features with the smallest number n of similar features to form the feature subset F.

[0056] Step 3: Perform a mutation operation on the feature subset F. Select a certain number mut num of features from the redundant similar feature set SF of the feature subset F and add them to the feature subset F to obtain the mutated feature subset Fm. Compare the classification effects of Fm and F. If the classification effect of Fm is better than that of F, let Fm be the final feature subset C in this round of iteration; otherwise, let F be the final feature subset C in this round of iteration. After T iterations, the target feature subset C1 is obtained.

[0057] From Figure 3From the comparison graph of experimental results, it can be seen that the present embodiment can effectively reduce the SRM features. In the embedding operation of 0.2bpp on the BOSSBase1.01 dataset using the S-UNIWARD steganography algorithm, the present embodiment can achieve a reduction of approximately 85% for the SRM features; in the embedding operation of 0.4bpp on the BOSSBase1.01 dataset using the S-UNIWARD steganography algorithm, the present embodiment can achieve a reduction of approximately 83% for the SRM features; in the embedding operation of 0.5bpp on the BOSSBase1.01 dataset using the S-UNIWARD steganography algorithm, the present embodiment can achieve a reduction of approximately 80% for the SRM features; in the embedding operation of 0.8bpp on the BOSSBase1.01 dataset using the S-UNIWARD steganography algorithm, the present embodiment can achieve a reduction of approximately 75% for the SRM features; in the embedding operation of 1.0bpp on the BOSSBase1.01 dataset using the S-UNIWARD steganography algorithm, the present embodiment can achieve a reduction of approximately 93% for the SRM features. From the experimental results, it can be seen that the method of the present invention can achieve feature reduction for the high-dimensional steganalysis feature set, which can greatly reduce the feature dimension and ensure that the detection accuracy is not affected.

Claims

1. A feature selection method for steganalysis based on the clonal selection algorithm, characterized in that Reduce the high-dimensional steganalysis feature set based on the clonal selection algorithm. By screening out the redundant features in the high-dimensional feature set, the purpose of reducing the feature dimension is achieved. Calculate the affinity value aff of the features by designing the affinity function Af of the features, and sort the features according to the affinity value. Select the optimal feature subset F from the features in a specific niche through T rounds of iteration, and perform a mutation operation on the feature subset F, that is, randomly select a certain number of similar features to be mutated and add them to the feature subset F to obtain the mutated feature subset Fm; compare the classification performance of the feature subset F and the mutated feature subset Fm to obtain the final feature subset C of this niche; obtain the target feature subset C1 after T rounds of iteration; including the following steps: Step 1, use the feature extraction function E to extract the steganalysis features of the carrier image dataset X and the stego-image dataset Y as the feature subset to be reduced Fc and the feature subset to be reduced Fs, and calculate the affinity values of each dimension of features in these two feature subsets to be reduced: First, divide the feature subset to be reduced into n {n ∈ N +}, calculate the affinity value for the features in each part; perform the maximum-minimum normalization operation on each dimension of features, and calculate the average value aver fc and aver fs in the corresponding carrier image or stego-image class for each dimension of features, and the standard deviation std fc and std fs , and finally calculate the affinity value aff of the feature according to the calculated average value and standard deviation: Where aff is the affinity value of each dimension of features in each feature subset, and finally obtain the set of feature affinity values of each feature subset to be reduced; Step 2, according to the feature affinity value aff obtained in step 1, sort each feature subset to be reduced according to its value, perform niche classification on the sorted features, and determine the features and the number of iterations T in each niche; the number of iterations T is dynamically determined according to the feature affinity value aff: where aff b and aff l respectively represent the maximum and minimum of the characteristic affinity of each part in n {n ∈ N +}, and niche refers to the iteration radius; Step 3, perform feature selection on the features in each round of iteration; first calculate the similarity between features, and the similarity between feature k and feature g is: For the calculated dist kg , find all features that are less than the similarity threshold S t {0 < S t < 1} as the similar features of feature k, and count the number of similar features for each feature; select the feature with the fewest number of similar features, and use its similar features as redundant features to form the redundant similar feature set SF until the redundant similar feature set SF is the same as all features in a specific niche; finally, output the feature subset F; Step 4, perform feature mutation on the feature subset F obtained in step 3, randomly select a certain number of features to be mutated from the redundant similar feature set SF and add them to the feature subset F to obtain the mutated feature subset Fm; the number of mutations is: mut num = min{P m * aver F , len(F s )}; where P m is the pre-set number of feature variations, which can be dynamically adjusted according to the feature dimension or affinity value, and aver F is the average value of the affinity values of all features in the feature subset F; len(F s ) is the number of features in the redundant similar feature set SF of F; the number mut num of the selected variant features is set to P m *aver F and the minimum value in len(F s ); Step 5, according to the feature subset F obtained in step 3 and the mutated feature subset Fm obtained in step 4, by comparing the detection accuracy of F and Fm in the integrated classifier M, take the larger value as the final feature subset C of this round of iteration; Step 6, according to the number of iterations T obtained in step 2, if still in the iteration process, continue to screen the features, repeat steps 3 to 5, and obtain the target feature subset C1 after T rounds of iteration.

2. The feature selection method for steganalysis based on the clonal selection algorithm according to claim 1, wherein The number of iterations T in step 2 can be dynamically adjusted according to the feature dimension and the affinity value.

3. The steganography analysis feature selection method based on the clonal selection algorithm according to claim 1, characterized in that The similarity dist in step 3 kg is dynamically adjusted according to the setting of n {n ∈ N +}}, and the similarity threshold S t {0 < S t < 1} can also be dynamically adjusted according to specific requirements.

4. A steganalysis feature selection method based on the clonal selection algorithm according to claim 1, characterized in that The number of mutations mut in step 4 num is dynamically adjusted according to the preset P m and the affinity value of the feature.

Images