Video steganalysis method based on motion vector domain of Skipped macroblock

Abstract

The present invention discloses a motion vector domain video steganalysis method based on skipped macroblocks, comprising: extracting first-category features of the skipped macroblocks by employing a recompression calibration method to obtain predicted motion vector recovery features of the skipped macroblocks; extracting second-category features of the skipped macroblocks by employing a recompression calibration method to obtain state transition probability features of the skipped macroblocks; and merging features by merging the predicted motion vector recovery features and state transition probability features of the skipped macroblocks to obtain comprehensive video steganalysis features of the skipped macroblocks. The motion vector domain video steganalysis method based on skipped macroblocks of the present invention extracts and merges predicted motion vector recovery features and state transition probability features of the skipped macroblocks through recompression calibration to improve the correct detection rate of steganalysis.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a motion vector domain video steganalysis method based on Skipped macroblocks. Background Art

[0002] Video steganalysis is the enemy of steganography. Its primary purpose is to determine whether video media contains secret information through statistical analysis. Specifically, the basic process of motion vector-based video steganalysis involves first decoding the compressed video stream, extracting statistical features related to motion vector modifications from the decoded parameters, then training a classifier, and finally classifying the detected objects to obtain the discrimination results. Due to the complexity of video coding, motion vector-based video steganography can cause different types of perturbations in the original bitstream's encoding parameters, thus enabling the extraction of video steganalysis features from different perspectives. Based on the starting point of feature extraction, video steganalysis based on motion vectors can be divided into five categories: the first category is methods based on the spatiotemporal statistical characteristics of motion vectors, because there is a strong correlation between motion vectors similar to that between pixels and / or DCT coefficients in an image; the second category is methods based on motion vector calibration, because the motion vectors of the steganographic video tend to be restored to the original motion vectors after calibration, thus providing a basis for steganalysis; the third category is methods based on the local optimality of motion vectors. Since the motion vector is a local optimal output process in the sense of rate distortion, the steganographic operation is likely to destroy this local optimality; the fourth category is a steganalysis algorithm designed based on the fact that the motion vectors of sub-blocks in a macroblock are usually inconsistent. It is the best performing method at present and can simultaneously detect steganography methods based on inter-frame prediction mode and motion vector domain; the fifth category is a motion vector domain video steganalysis method based on convolutional neural network and skipped macroblock, which is currently in the initial stage of research.

[0003] As can be seen from the current state of steganalysis research, since motion vector domain steganography uses motion vectors as its carrier, all motion vector domain steganalysis algorithms directly extract features from motion vectors. However, this ignores the fact that the video compression and encoding process is a closely interconnected process. Modifying the motion vector not only perturbs its own statistical characteristics but can also cause anomalies in the statistical characteristics of other coding parameters. In normal video compression and encoding, for example, inter-frame coded P frames in the H.264 / AVC standard primarily consist of P macroblocks and P-Skip macroblocks. Macroblocks can be divided into P-Skip macroblocks and P macroblocks of varying sizes based on different macroblock partitioning schemes. Each P macroblock contains a set of motion vectors (horizontal and vertical components) that point to the optimal reference block. P-Skip macroblocks do not directly contain motion vectors; their optimal reference block is determined by a predicted motion vector. The encoder calculates its predicted motion vector (MVP) based on the motion vectors of the three coded macroblocks in the P-Skip block's neighborhood. Motion vector domain steganography directly modifies the P macroblock's motion vector, causing changes in the P macroblock's motion vector statistics. This is the starting point for all current motion vector domain-based video steganalysis feature extraction. Although the P-Skip macroblock does not have a motion vector directly used for steganographic embedding, its predicted motion vector is determined by the motion vectors of the three encoded macroblocks in its neighborhood. If the encoded macroblocks in the neighborhood of the P-Skip macroblock are disturbed during the steganography process, their predicted motion vectors may also be passively modified, causing the best matching block corresponding to the P-Skip macroblock to change from optimal to suboptimal. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a motion vector domain video steganalysis method based on Skipped macroblocks to improve the correct detection rate of video steganalysis.

[0005] In order to solve the above technical problems, the purpose of the present invention is achieved through the following technical solutions: providing a motion vector domain video steganalysis method based on Skipped macroblocks, including: extraction of the first type of features of the Skipped macroblock: adopting a heavy compression calibration method to design and obtain the predicted motion vector recovery features of the Skipped macroblock; extraction of the second type of features of the Skipped macroblock: adopting a heavy compression calibration method to design and obtain the state transition probability features of the Skipped macroblock; feature merging: merging the predicted motion vector recovery features and state transition probability features of the Skipped macroblock to obtain the comprehensive video steganalysis features of the Skipped macroblock.

[0006] The beneficial technical effect of the present invention is that the motion vector domain video steganalysis method based on the Skipped macroblock of the present invention extracts the predicted motion vector recovery features and state transition probability features of the Skipped macroblock through recompression calibration, and merges the two types of extracted features to obtain the comprehensive video steganalysis features of the Skipped macroblock, so as to improve the comprehensive detection capability, thereby improving the correct detection rate of the steganalysis. Moreover, the analysis and extraction based on the Skipped macroblock is insensitive to different steganography algorithms, so that the corresponding comprehensive video steganalysis features obtained by extraction have stronger adaptability, good detection performance, and are less affected by the encoding parameters of the recompression calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0008] Figure 1 A schematic diagram of the process of a motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention;

[0009] Figure 2 A schematic diagram of the first sub-process of the motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention;

[0010] Figure 3 A schematic diagram of a second flow chart of a motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention;

[0011] Figure 4 The statistical distribution diagram of the predicted motion vector difference operator of the P-Skip macroblock in the carrier video and the steganographic video after recompression calibration;

[0012] Figure 5 Schematic diagram of the distribution of macroblock partitioning methods for macroblocks that are P-Skip blocks after recompression calibration in a video bitstream;

[0013] Figure 6 A schematic diagram showing experimental comparison results of the motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention and the AoSO, NPELO, and MVC steganalysis methods for the correct extraction rates of the three different steganographic methods at different compression quality factors;

[0014] Figure 7Schematic diagram of experimental results of recompression calibration of the motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention using a compression quality factor different from that used for the video stream. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0016] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0017] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0018] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0019] See also Figure 1 As shown, Figure 1 A schematic flow chart of a motion vector domain video steganalysis method based on Skipped macroblocks provided by an embodiment of the present invention includes the following steps:

[0020] Step S11, extracting the first type of features of the Skipped macroblock: adopting a recompression calibration method to obtain the predicted motion vector recovery features of the Skipped macroblock;

[0021] Step S12, extracting the second type of features of the Skipped macroblock: adopting a recompression calibration method to design and obtain the state transition probability features of the Skipped macroblock;

[0022] Step S13, feature merging: merging the predicted motion vector recovery feature and the state transition probability feature of the Skipped macroblock to obtain the comprehensive video steganalysis feature of the Skipped macroblock.

[0023] The skipped macroblock can be a P-Skip macroblock in an inter-frame coded P-frame. The calibration concept of the recompression calibration method is derived from JPEG image steganalysis. It states that after recompression and calibration, the encoding parameters of a JPEG image can be restored to their original state to a certain extent. After recompression and calibration of the stegographed video using the same parameters, its motion vector will exhibit the property of returning to its original value. The motion vector domain video steganalysis method based on skipped macroblocks extracts predicted motion vector recovery features and state transition probability features from the skipped macroblocks through recompression and calibration. These two extracted features are then merged to obtain comprehensive video steganalysis features for the skipped macroblocks, thereby improving comprehensive detection capabilities and thus the correct detection rate of steganalysis. Furthermore, the analysis and extraction based on skipped macroblocks is insensitive to different steganographic algorithms, resulting in a more adaptable and better detection performance of the corresponding comprehensive video steganalysis features, which are less affected by the recompression and calibration encoding parameters.

[0024] Combine Figure 2 Specifically, the recompression calibration method in step S11 and step S12 includes the following steps:

[0025] Step S101, video stream decoding: decode the video stream to obtain a decoded spatial domain YUV file; wherein the video stream standard may be H.264 / AVC, and the YUV file is a picture file consisting of three parts: Y, U, and V, where Y represents brightness, and U and V both represent color chrominance.

[0026] Step S102: Coding parameter extraction: Extract the coding parameter information of the video stream. The coding parameter information includes the frame rate, resolution, GOP structure, compression quality factor (QP), bit rate, macroblock partitioning method, and motion vectors. The coding parameter information is extracted during the decoding process of the video stream. The coding parameter information at this time is the parameter information obtained during the initial encoding of the video stream.

[0027] Step S103, re-encoding: using an encoder to re-encode the spatial domain YUV file obtained after decoding according to the extracted encoding parameter information to obtain a secondary compressed code stream.

[0028] Step S104: Decoding the secondary compressed bitstream: Decode the secondary compressed bitstream to obtain the macroblock partitioning method and motion vectors from the encoding parameter information of the secondary compressed bitstream. The encoding parameter information is extracted during the decoding of the video bitstream. The encoding parameter information at this time is the parameter information obtained when the secondary compressed bitstream was re-encoded.

[0029] Step S105: Video steganalysis feature extraction: Video steganalysis features are extracted based on the encoding parameter information of the video bitstream and the macroblock partitioning method and motion vectors in the encoding parameter information of the secondary compressed bitstream. The extracted video steganalysis features are features that can effectively distinguish between the carrier video and the stego video, thereby achieving the purpose of detecting steganalysis. For P frames, in the H.264 / AVC standard, a macroblock must meet the following conditions to be encoded as a P-Skip macroblock:

[0030] The optimal motion-compensated macroblock size is 16×16; the reference frame must be the previous frame; the motion vector and the predicted motion vector are the same; and the macroblock is an all-zero block after pixel residual transformation.

[0031] Combine Figure 3 Specifically, the step of extracting video steganalysis features in step S105 is as follows:

[0032] Step S1051: setting a plurality of consecutive P frames as a feature extraction window to group the video stream and the secondary compression stream respectively;

[0033] Step S1052: extracting video steganalysis features from the corresponding grouped video stream and secondary compressed stream.

[0034] Preferably, the step S1051 is specifically as follows:

[0035] A plurality of non-overlapping continuous P frames are used as a feature extraction window to group the video stream and the secondary compression stream respectively, to obtain a plurality of video stream extraction groups and a plurality of secondary compression stream extraction groups, wherein the video stream extraction groups and the secondary compression stream extraction groups correspond one to one.

[0036] Specifically, when designing to obtain the predicted motion vector recovery feature of the Skipped macroblock, the step S1052 is:

[0037] The secondary compression code stream extraction group is searched for a Skipped macroblock in which the macroblock division method remains unchanged before and after the recompression calibration in the corresponding video code stream extraction group, and the predicted motion vector recovery feature of the searched Skipped macroblock is extracted.

[0038] Among them, the predicted motion vector recovery feature of the Skipped macroblock is a 5-dimensional feature, the state transition probability feature of the Skipped macroblock is a 6-dimensional feature, and the comprehensive video steganalysis feature of the Skipped macroblock obtained after merging is an 11-dimensional feature. Figure 4The statistical distribution of the predicted motion vector difference operator of the P-Skip macroblocks in the carrier video and the stego video after recompression and calibration is shown, and only those macroblocks that are P-Skip macroblocks in the original video and are also P-Skip macroblocks in the calibrated video are counted. Figure 4 As shown, the horizontal axis is the value of the predicted motion vector differential operator before and after the recompression calibration of the P-Skip macroblock, and the vertical axis is the empirical probability of the value of the predicted motion vector differential operator appearing in all P-Skip macroblocks. The predicted motion vector differential operator before and after the recompression calibration of the P-Skip macroblock in the carrier video (Cover) is mainly 0, accounting for 85% when the compression quality factor (QP) is 25, and accounting for 83.1% when the compression quality factor (QP) is 35. There are very few P-Skip macroblocks whose predicted motion vector differential operator value before and after the recompression calibration exceeds 2. Therefore, the predicted motion vector of the P-Skip macroblock can be effectively and accurately restored through recompression calibration. Secondly, regardless of whether the compression quality factor (QP) is 25 or 35, compared with the carrier video (Cover), the probability of P-Skip macroblocks with the value of the predicted motion vector differential operator being 0 before and after recompression calibration in the steganographic video (Stego) has decreased significantly, while the probability of P-Skip macroblocks with the value of the predicted motion vector differential operator being 1 before and after recompression calibration has increased significantly. The probability of P-Skip macroblocks with the value of the predicted motion vector differential operator being greater than 1 before and after recompression calibration has also increased to a certain extent.

[0039] Specifically, it is assumed that there are multiple P-Skip macroblocks in the video stream and the macroblock division method remains unchanged before and after recompression calibration, that is, it is assumed that there are multiple macroblocks in the video stream that are all Skipped macroblocks before and after recompression calibration. Then, the predicted motion vector recovery feature of the Skipped macroblock can be expressed by formula (1):

[0040]

[0041] Among them, B i Indicates the i-th Skipped macroblock whose macroblock division method remains unchanged before and after recompression calibration, n represents the total number of Skipped macroblocks whose macroblock division method remains unchanged before and after recompression calibration, diff represents the predicted motion vector difference operator before and after recompression calibration of Skipped macroblocks to describe the degree of change of the predicted motion vector before and after recompression calibration of Skipped macroblocks, diff Birepresents the differential operator of the predicted motion vector before and after the recompression calibration of the i-th Skipped macroblock whose macroblock division method remains unchanged before and after the recompression calibration, f1(·) represents the predicted motion vector recovery feature of the Skipped macroblock, k1 represents the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock, and the number of dimensions of the predicted motion vector recovery feature of the Skipped macroblock is 5, then the value of the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock can be any natural number from 0 to 4, the value of the predicted motion vector differential operator before and after the recompression calibration of the Skipped macroblock can correspond one-to-one to the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock, Pr(·) represents the probability function, Pr(diff=k1) represents the probability that the value (diff) of the predicted motion vector differential operator before and after the recompression calibration of the Skipped macroblock is equal to the value (k1) of the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock, Represents the predicted motion vector difference distribution function of the Skipped macroblock whose macroblock division mode remains unchanged before and after recompression calibration, when the value of the difference operator of the predicted motion vector of the Skipped macroblock whose macroblock division mode remains unchanged before and after recompression calibration (diff Bi ) is equal to the value (k1) of the dimension number of the predicted motion vector recovery feature of the Skipped macroblock, then The value of is 1, otherwise, The value is 0.

[0042] Specifically, the difference operator can be calculated using formula (2):

[0043] diff=|hh′|+|vv′| (2)

[0044] Among them, diff represents the predicted motion vector differential operator before and after the Skipped macroblock is recompressed and calibrated, h is the horizontal component of the motion vector before the Skipped macroblock is recompressed and calibrated, h′ is the horizontal component of the motion vector after the Skipped macroblock is recompressed and calibrated, v is the vertical component of the motion vector before the Skipped macroblock is recompressed and calibrated, and v′ is the vertical component of the motion vector after the Skipped macroblock is recompressed and calibrated.

[0045] Specifically, when designing to obtain the state transition probability feature of the Skipped macroblock, the step S1052 is:

[0046] The video code stream extraction group is searched for Skipped macroblocks in the corresponding secondary compression code stream extraction group, whose macroblock division mode remains unchanged before and after recompression and calibration, and the state transition probability feature of the searched Skipped macroblock is extracted.

[0047] in, Figure 5 The distribution of macroblock partitioning methods of macroblocks that are P-Skip blocks after recompression calibration in the video bitstream is shown, such as Figure 5 As shown in the figure, when the compression quality factor (QP) is 25, for the carrier video (Cover), the macroblocks that are P-Skip macroblocks in the secondary compression bitstream are primarily divided into P-Skip macroblocks and 16×16 P macroblocks in the video bitstream, with corresponding proportions of 68.4% and 21.4%, respectively. For the steganographic video (Stego), the macroblocks that are P-Skip macroblocks in the secondary compression bitstream are also primarily divided into P-Skip macroblocks and 16×16 P macroblocks in the video bitstream, but with corresponding proportions of 42.7% and 43.3%, respectively. After recompression calibration, fewer P-Skip macroblocks and more 16×16 P macroblocks are encoded as P-Skip macroblocks. The situation when the compression quality factor (QP) is 35 is similar to that when the compression quality factor (QP) is 25, and will not be repeated here.

[0048] Specifically, the set of macroblock partitioning methods of the Skipped macroblock in the secondary compression code stream in the video code stream is set to Partition, Partition = {P-Skip, 16×16, 16×8, 8×16, 8×8, else}, and the number of Skipped macroblocks in the secondary compression code stream is set to m. Then, the state transition probability characteristics of the Skipped macroblock can be expressed by formula (3):

[0049]

[0050] Among them, C jRepresents the jth Skipped macroblock in the secondary compression code stream, m represents the total number of Skipped macroblocks in the secondary compression code stream, k2 represents the dimensional sequence number of the state transition probability feature of the Skipped macroblock, and the total number of dimensions of the comprehensive video steganalysis feature of the Skipped macroblock is 11, which is the sum of the number of dimensions of the predicted motion vector recovery feature of the Skipped macroblock and the number of dimensions of the state transition probability feature of the Skipped macroblock. The value of the dimensional sequence number of the comprehensive video steganalysis feature of the Skipped macroblock can be any natural number from 0 to 10, the number of dimensions of the state transition probability feature of the Skipped macroblock is 6, the value of the dimensional sequence number of the state transition probability feature of the Skipped macroblock can be any natural number from 5 to 10, f2(·) represents the state transition probability feature of the Skipped macroblock, partition(k2-5) represents the Skipped in the secondary compression code stream. The macroblock partitioning mode of the macroblock in the video code stream, k2-5 represents the index number of the element in the set Partition, then the macroblock partitioning mode of the Skipped macroblock in the secondary compression code stream corresponding to partition(0) in the video code stream is the macroblock partitioning mode corresponding to the first element in the set Partition, which is P-Skip in this embodiment, and so on; Pr(·) represents the probability function, Pr(partition(k2-5)) represents the probability of the macroblock partitioning mode of the Skipped macroblock in the secondary compression code stream in the video code stream, φ(x,y) represents the function for determining the macroblock partitioning mode of the macroblock, x represents the macroblock, y represents the macroblock partitioning mode, when the macroblock partitioning mode of macroblock x is equal to the macroblock partitioning mode y, then macroblock x adopts macroblock partitioning mode y, and φ(x,y) takes the value of 1; otherwise, macroblock x does not adopt macroblock partitioning mode y, and φ(x,y) takes the value of 0. When k2 = 5, the state transition probability feature of the skipped macroblock is the probability that the skipped macroblock in the secondary compression bitstream is also a skipped macroblock in the video bitstream. The state transition probability feature of the skipped macroblock is a six-dimensional feature, corresponding to the probabilities that the skipped macroblock in the secondary compression bitstream is a P-Skip macroblock, a 16×16 P macroblock, a 16×8 P macroblock, an 8×16 P macroblock, an 8×8 P macroblock, and P macroblocks of other sizes in the video bitstream.

[0051] Specifically, the following table shows the correct detection rates of the motion vector domain video steganalysis method based on skipped macroblocks for three different steganographic algorithms at different compression quality factors (QPs) and different embedding capacities. As shown in the table, at an extremely low embedding capacity (0.05 bpnsmv), the correct detection rates of the motion vector domain video steganalysis method based on skipped macroblocks for the three steganographic algorithms averaged 72.83%, 74.88%, and 69.48%, respectively. This is a relatively high detection rate, demonstrating that the motion vector domain video steganalysis method based on skipped macroblocks performs well at low embedding capacities. While the correct detection rate improves with increasing embedding capacity, the increase is modest. This is because P-Skip macroblocks typically appear clustered in P frames. Once the predicted motion vectors of one or more macroblocks are disturbed, the other macroblocks are likely to be disturbed as well. However, as the embedding capacity increases, despite the increase in the number of disturbed motion vectors, the probability of P-Skip macroblocks being disturbed does not increase significantly. Secondly, the predicted motion vector recovery and state transition probability features of skipped macroblocks show little difference in detection rate across different steganographic algorithms. This is primarily because different steganographic algorithms embed messages in different macroblocks (subblocks) with motion vectors. However, the predicted motion vector recovery and state transition probability features of skipped macroblocks are not directly extracted from these macroblocks with motion vectors. Instead, they are indirectly extracted from the predicted motion vectors of skipped macroblocks (i.e., P-Skip macroblocks). Therefore, they are insensitive to different steganographic algorithms. This also indicates that the predicted motion vector recovery and state transition probability features of skipped macroblocks have greater adaptability. Finally, for different compression rates, the correct detection rate of the predicted motion vector recovery and state transition probability features of skipped macroblocks is higher when the compression quality factor (QP) is 25 than when the compression quality factors (QP) are 15 and 35. This is because when the compression quality factor (QP) is 15, the macroblock division is finer, and fewer macroblocks are divided into P-Skip macroblocks. Therefore, fewer P-Skip macroblocks can be used to extract the predicted motion vector recovery features and state transition probability features of the skipped macroblocks. When the compression quality factor (QP) is 35, the video compression ratio is higher, and a large portion of the macroblocks in the P frame are divided into P-Skip macroblocks, while the number of ordinary P macroblocks is small. Therefore, the number of motion vector carriers used for steganography is also smaller. At the same relative embedding capacity, the motion vector is less perturbed by steganography. Therefore, the steganalysis detection performance is actually reduced compared to the compression quality factor (QP) of 25.

[0052]

[0053] See Figure 6 , Figure 6 The experimental comparison results of the correct extraction rate of the motion vector domain video steganalysis method based on Skipped macroblocks of the present invention and the AoSO, NPELO and MVC steganalysis methods at different compression quality factors (QP) are shown. The motion vector domain video steganalysis method based on Skipped macroblocks is denoted as Proposed SMCF. Figure 6 As shown, the compression quality factor (QP) between Figure (a) and Figure (b) is different. Overall, the motion vector domain video steganalysis method based on skipped macroblocks significantly outperforms AoSO and NPELO under different conditions, indicating that the motion vector domain video steganalysis method based on skipped macroblocks has better adaptability and detection performance. As the embedding capacity increases, the improvement rate of the correct detection rate of the three steganalysis methods, AoSO, NPELO, and MVC, is better than that of the motion vector domain video steganalysis method based on skipped macroblocks. This is because the three methods used for comparison all extract features directly in the motion vector domain. As the embedding capacity increases, the disturbance caused by the steganographic operation to the motion vector also increases, making it easier to detect at large capacity. At low embedding capacity (i.e., 0.05, 0.1, and 0.2 bpnsmv), the detection effect of the motion vector domain video steganalysis method based on Skipped macroblocks is better than that of MVC. However, at high embedding capacity (i.e., 0.3 and 0.4 bpnsmv), the detection effect of the motion vector domain video steganalysis method based on Skipped macroblocks is equal to or lower than that of MVC. Figure 6As shown, when the embedding capacity is less than 0.3 bpnsmv, the correct detection rate of the motion vector domain video steganalysis method based on skipped macroblocks is superior to that of MVC. However, when the embedding capacity is between 0.3 bpnsmv and 0.4 bpnsmv, MVC outperforms the motion vector domain video steganalysis method based on skipped macroblocks. This is because when the embedding capacity is larger, MVC can better detect the perturbations to the consistency of intra-block motion vectors in the MVPLO algorithm. The detection rates of the four steganalysis methods at a compression quality factor (QP) of 35 are slightly lower than those at a QP of 25, but the degradation of the features of the motion vector domain video steganalysis method based on skipped macroblocks is smaller. This is because at a higher compression quality factor (i.e., a QP of 35), most macroblocks are divided into P-Skip macroblocks, leaving fewer ordinary P macroblocks and, consequently, fewer motion vector carriers used for steganography. Therefore, at the same relative embedding capacity, the perturbations to the motion vectors are also smaller. For steganalysis methods such as AoSO, NPELO, and MVC that directly extract information features in the motion vector domain, the correct detection rate will decrease under high quality factors. However, the motion vector domain video steganalysis method based on Skipped macroblocks of the present invention is less affected by changes in the compression factor because it does not directly extract features from motion vectors.

[0054] Specifically, the following table shows the experimental results of the effects of the comprehensive video steganalysis features of the Skipped macroblock (denoted as SMCF features) obtained by combining the predicted motion vector recovery features and state transition probability features of the macroblock in the motion vector domain video steganalysis method based on the Skipped macroblock of the present invention, the predicted motion vector recovery features of the individual Skipped macroblock (denoted as SMCF-part1 features), and the state transition probability features of the individual Skipped macroblock (denoted as SMCF-part2 features) on the detection performance of the PCAMV steganography algorithm. As shown in the table, according to the data of the experimental results, the detection performance of the comprehensive video steganalysis features of the Skipped macroblock is better than that of the Skipped macroblock and its subset. The predicted motion vector recovery feature of the block and the state transition probability feature of the Skipped macroblock show that there is no obvious conflict between the predicted motion vector recovery feature and the state transition probability feature of the Skipped macroblock, and the combination of the two can effectively improve the comprehensive detection capability. When the compression quality factor (QP) is 25, the detection capabilities of the predicted motion vector recovery feature and the state transition probability feature of the Skipped macroblock are respectively 15% and 5% lower than the comprehensive video steganalysis feature of the Skipped macroblock on average, indicating that the state transition probability feature of the Skipped macroblock plays a more important role, which means that the impact of the steganographic operation on the state of the Skipped macroblock is greater than the impact of the steganographic operation on the predicted motion vector of the Skipped macroblock.

[0055]

[0056]

[0057] See Figure 7 , Figure 7 The experimental results of recompression calibration using a compression quality factor different from that used for video streams are shown, where the steganography algorithm is the PCAMV steganography algorithm, such as Figure 7As shown in the figure, in the experiment, the compression quality factor (QP) of the original carrier video and the stego video is 25. During the recompression calibration, the compression quality factor (QP) is set to 25 and 22 and 28 which are close to 25. When the compression quality factor (QP) of the recompression calibration is the same as the compression quality factor (QP) of the original carrier video and the stego video, the detection performance of the comprehensive video steganalysis feature of the Skipped macroblock is the highest. Although when the compression quality factor (QP) of the recompression calibration is different from the compression quality factor (QP) of the original carrier video and the stego video, the comprehensive video steganalysis feature of the Skipped macroblock is mainly The predicted motion vector recovery feature and state transition probability feature of the Skipped macroblock are combined to obtain the feature. Although the compression quality factor (QP) affects the number of Skipped macroblocks, the larger the compression quality factor (QP), the more Skipped macroblocks there are. However, the compression quality factor (QP) has little effect on the distribution of Skipped macroblocks. Therefore, even if the compression quality factor (QP) has a certain degree of mismatch and the detection performance of the comprehensive video steganalysis feature of the Skipped macroblock is reduced, the reduction is small, and the comprehensive video steganalysis feature of the Skipped macroblock can still maintain a certain applicability.

[0058] Specifically, the following table shows the average time required to extract four different video steganalysis features from a single video sequence at different compression quality factors (QPs). The time unit is seconds, the video sequence format is CIF, and the video sequence is 240 frames. The dimensions of the AoSO feature, NPELO feature, MVC feature, and SMCF feature (comprehensive video steganalysis feature for skipped macroblocks) are 18, 36, 12, and 11, respectively. As shown in the table, the MVC feature has the highest performance, because it primarily extracts motion vectors within a macroblock or subblock and calculates their correlation. Secondly, the computational complexity of the AoSO and NPELO features is relatively similar, but significantly higher than that of the MVC feature. This is because both AoSO and NPELO require traversing the neighborhood of motion vector values to calculate their local optimality. The motion vector domain video steganalysis method based on skipped macroblocks has the highest time complexity. This is because it is implemented based on recompression and calibration, and the runtime is primarily concentrated in two video decoding operations and one video recompression and calibration operation. Furthermore, the performance of the AoSO, NPELO, and MVC features improves with increasing compression quality factor (QP). This is because a larger QP results in smaller motion vectors in the bitstream, reducing the amount of data required for processing. However, the performance of the SMCF feature (Skipped Macroblock Comprehensive Video Steganalysis Feature) is not significantly different at different compression quality factors (QP) because it does not directly extract features from the motion vectors of macroblocks or sub-blocks.

[0059]

[0060] In summary, the motion vector domain video steganalysis method based on Skipped macroblocks of the present invention extracts the predicted motion vector recovery features and state transition probability features of Skipped macroblocks through recompression calibration, and merges the two types of extracted features to obtain the comprehensive video steganalysis features of the Skipped macroblocks, so as to improve the comprehensive detection capability, thereby improving the correct detection rate of steganalysis. Moreover, the analysis and extraction based on Skipped macroblocks is insensitive to different steganography algorithms, so that the corresponding comprehensive video steganalysis features obtained by extraction have stronger adaptability, good detection performance, and are less affected by the encoding parameters of recompression calibration.

[0061] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A motion vector domain video steganalysis method based on Skipped macroblocks, characterized by: include: Extraction of the first type of features of Skipped macroblocks: Using the recompression calibration method to obtain the predicted motion vector recovery features of the Skipped macroblocks; Extraction of the second type of features of the Skipped macroblock: using a recompression calibration method to obtain the state transition probability features of the Skipped macroblock; feature Merging: merging the predicted motion vector recovery feature and the state transition probability feature of the Skipped macroblock to obtain the comprehensive video steganalysis feature of the Skipped macroblock; The recompression calibration method for extracting the first type of features of the Skipped macroblock and the second type of features of the Skipped macroblock specifically includes the following steps: Video stream decoding: decode the video stream and obtain the decoded spatial domain YUV file; Coding parameter extraction: Extracts the coding parameter information of the video stream, including frame rate, resolution, GOP structure, compression quality factor (QP), bit rate, macroblock division method and motion vector; Re-encoding: The decoded spatial domain YUV file is re-encoded using the encoder according to the extracted encoding parameter information to obtain a secondary compressed code stream; Secondary compression code stream decoding: Decode the secondary compression code stream to obtain the macroblock division mode and motion vector in the encoding parameter information of the secondary compression code stream; Video steganalysis feature extraction: Analyze the macroblock division method and motion vector in the encoding parameter information of the video bitstream and the secondary compression bitstream to extract video steganalysis features.

2. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 1, characterized in that: The step of extracting video steganalysis features in the video steganalysis feature extraction is specifically as follows: Set a number of consecutive P frames as a feature extraction window to group the video stream and the secondary compression stream respectively; Extract video steganalysis features from the corresponding grouped video stream and secondary compressed stream.

3. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 2, characterized in that: The steps of setting a plurality of consecutive P frames as a feature extraction window to group the video stream and the secondary compression stream are specifically as follows: A plurality of non-overlapping continuous P frames are used as a feature extraction window to group the video stream and the secondary compression stream respectively, to obtain a plurality of video stream extraction groups and a plurality of secondary compression stream extraction groups, wherein the video stream extraction groups and the secondary compression stream extraction groups correspond one to one.

4. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 3, characterized in that: When designing to obtain the predicted motion vector recovery features of the Skipped macroblock, the steps of extracting the video steganalysis features from the corresponding grouped video stream and the secondary compressed stream are specifically as follows: The secondary compression code stream extraction group is searched for a Skipped macroblock in which the macroblock division method remains unchanged before and after the recompression calibration in the corresponding video code stream extraction group, and the predicted motion vector recovery feature of the searched Skipped macroblock is extracted.

5. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 4, characterized in that: The predicted motion vector recovery feature of the Skipped macroblock is expressed by the following formula: Among them, B i Indicates the i-th Skipped macroblock whose macroblock division method remains unchanged before and after recompression calibration, n represents the total number of Skipped macroblocks whose macroblock division method remains unchanged before and after recompression calibration, diff represents the predicted motion vector difference operator before and after recompression calibration of Skipped macroblocks, represents the differential operator of the predicted motion vector before and after recompression calibration of the i-th Skipped macroblock whose macroblock division method remains unchanged before and after recompression calibration, f1(·) represents the predicted motion vector recovery feature of the Skipped macroblock, k1 represents the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock, and the value of the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock is any natural number from 0 to 4. Pr(·) represents the probability function, Pr(diff=k1) represents the probability that the value of the predicted motion vector differential operator before and after recompression calibration of the Skipped macroblock is equal to the value of the dimensional sequence number of the predicted motion vector recovery feature of the Skipped macroblock, Represents the predicted motion vector difference distribution function of the Skipped macroblock whose macroblock division mode remains unchanged before and after recompression calibration. When the value of the differential operator of the predicted motion vector of the i-th Skipped macroblock whose macroblock division mode remains unchanged before and after recompression calibration is equal to the value of the dimension sequence number of the predicted motion vector recovery feature of the Skipped macroblock, then The value of is 1, otherwise, The value is 0.

6. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 3, characterized in that: When designing to obtain the state transition probability features of the Skipped macroblock, the steps of extracting the video steganalysis features from the corresponding grouped video stream and the secondary compressed stream are specifically as follows: The video code stream extraction group is searched for Skipped macroblocks in the corresponding secondary compression code stream extraction group, whose macroblock division mode remains unchanged before and after recompression and calibration, and the state transition probability feature of the searched Skipped macroblock is extracted.

7. The motion vector domain video steganalysis method based on Skipped macroblocks according to claim 6, characterized in that: The state transition probability characteristic of the Skipped macroblock is expressed by the following formula: Among them, C j represents the jth Skipped macroblock in the secondary compression code stream, m represents the total number of Skipped macroblocks in the secondary compression code stream, k2 represents the dimension number of the state transition probability feature of the Skipped macroblock, the dimension number of the state transition probability feature of the Skipped macroblock is any natural number from 5 to 10, f2(·) represents the state transition probability feature of the Skipped macroblock, partition(k2-5) represents the macroblock partitioning method of the Skipped macroblock in the secondary compression code stream in the video code stream, k2-5 represents the index number of the element in the set Partition, then partition(0) corresponds to the Skipped macroblock in the secondary compression code stream. The macroblock partitioning mode of the skipped macroblock in the video bitstream is the macroblock partitioning mode corresponding to the first element in the set Partition, and so on; Pr(·) represents the probability function, Pr(partition(k2-5)) represents the probability of the macroblock partitioning mode of the skipped macroblock in the secondary compression bitstream in the video bitstream, φ(x,y) represents the function of judging the macroblock partitioning mode of the macroblock, x represents the macroblock, y represents the macroblock partitioning mode, when the macroblock partitioning mode of macroblock x is equal to the macroblock partitioning mode y, then the macroblock x adopts the macroblock partitioning mode y, and φ(x,y) takes the value of 1; otherwise, the macroblock x does not adopt the macroblock partitioning mode y, and φ(x,y) takes the value of 0.

Images