Disclosure of Invention
In view of the foregoing problems, an object of the present invention is to provide a general audio steganography analysis method based on a convolutional neural network and multitask learning, which can effectively improve the detection effect of an audio steganography algorithm and the detection capability of an unknown steganography algorithm.
In order to achieve the purpose, the technical scheme of the invention is as follows: a general audio steganalysis method based on a convolutional neural network and multitask learning is characterized in that: the network framework corresponding to the method comprises a feature extraction sub-network, a two-classification sub-network and a multi-classification sub-network, the method comprises,
s1, inputting audio data;
s2, outputting a steganalysis feature vector F through a feature extraction sub-network;
s3, judging whether the audio data is a steganographic carrier or not through a sub-classification network, if so, sequentially executing S4-S8, and if not, outputting the audio data as normal audio;
s4, steganalysis feature directionThe quantity F is output through a binary classification sub-network to obtain a binary steganography prediction probability vector
Computing binary steganographic prediction probability vectors
Two-element steganographic label vector y ═ y after One-hot coding
0,y
1]Cross entropy loss of L
m,
Wherein y is
i∈ {0,1}, i denotes the class index, i ∈ [0, 1]]Updating parameters of the two classification sub-networks through a back propagation error and gradient descent algorithm;
s5, outputting the steganography analysis feature vector F through a multi-classification sub-network to obtain the prediction probability value of the steganography algorithm type
Calculating a predicted probability value
Steganographic class tag m ═ m after encoding with One-hot
0,m
1,…,m
M-1]Cross entropy loss of L
a,
Wherein M represents the number of M different steganographic algorithms contained in the training set data, and sub-network parameters are classified according to the number through a back propagation error and gradient descent algorithm;
s6, according to the combined loss L ═ Lm+λLaSetting lambda as an auxiliary task weight factor;
s7, calculating a confidence value C (m) of the prediction probability through the multi-classification sub-network;
and S8, judging whether the confidence value C (m) is greater than a set experience threshold value CT, if so, outputting the result as an unknown steganography algorithm, and if not, outputting the type of the steganography algorithm.
Further, the feature extraction sub-network in S2 includes an audio preprocessing layer and 5 concatenated convolution groups after the audio preprocessing layer, that is, a 1 st convolution group, a 2 nd convolution group, a 3 rd convolution group, a 4 th convolution group, and a 5 th convolution group.
Further, the audio preprocessing layer is composed of 4 1 × 5 convolution kernels D1-D4, and the initial weights are respectively:
D1=[1,-1,0,0,0],D1=[1,-2,1,0,0],D1=[1,-3,3,1,0],D1=[1,-4,6,-4,1];
the 1 st convolution group includes a 1 × 1 first convolution layer, a 1 × 5 second convolution layer, and a 1 × 1 third convolution layer;
the 2 nd convolution group, the 3 rd convolution group, the 4 th convolution group and the 5 th convolution group respectively comprise a 1 x 5 convolution layer, a 1 x 1 convolution layer and a mean pooling layer, wherein the mean pooling layer of the 5 th convolution group is a global mean pooling layer;
the steganalysis feature vector is a 256-dimensional vector.
Furthermore, the audio preprocessing layer adopts a differential filtering design.
Further, the first convolution layer in the 1 st convolution group is activated by using a truncated linear unit TLU.
Further, the two classification subnetworks include a fully connected layer having 128 neurons and a binary steganographic label prediction layer.
Further, the multi-class subnetwork comprises two cascaded fully-connected layers and steganographic class label prediction layers, wherein the two cascaded layers respectively have 128 neurons and 64 neurons.
Further, the confidence value C (m) in S8 is calculated by
Setting an empirical threshold CT ═ 0.5 ═ c (m) max, where c (m)
max=logM。
Compared with the prior art, the invention has the advantages that:
by providing the audio universal steganography analysis model and the audio universal steganography analysis method based on the convolutional neural network and the multitask learning, the detection effect of various audio steganography algorithms is effectively improved; the method improves the detection capability of the unknown steganography algorithm, and is convenient for the application of the audio steganography analysis technology in the scene of obtaining evidence of big data of the complex internet.
Detailed Description
The following detailed description of embodiments of the invention is merely exemplary in nature and is intended to be illustrative of the invention and not to be construed as limiting the invention.
The invention discloses a general audio steganalysis method based on convolutional neural network and multitask learning, which comprises the following steps,
s1, inputting audio data;
s2, outputting a steganalysis feature vector F through a feature extraction sub-network;
s3, judging whether the audio data is a steganographic carrier or not through a sub-classification network, if so, sequentially executing S4-S8, and if not, outputting the audio data as normal audio;
s4, outputting the feature vector F of steganalysis through two classification sub-networks to obtain a binary steganalysis prediction probability vector
Computing binary steganographic prediction probability vectors
Binary steganographic label vector y ═ y after One-hot coding
0,y
1]Cross entropy loss of L
m,
Wherein y is
i∈ {0,1}, i denotes the class index, i ∈ [0, 1]]Updating parameters of the two classification sub-networks through a back propagation error and gradient descent algorithm;
s5, outputting the steganography analysis feature vector F through a multi-classification sub-network to obtain the prediction probability value of the steganography algorithm type
Calculating a predicted probability value
Steganographic class tag m ═ m after encoding with One-hot
0,m
1,…,m
M-1]Cross entropy loss of L
a,
Wherein M represents the number of M different steganographic algorithms contained in the training set data, and sub-network parameters are classified according to the number through a back propagation error and gradient descent algorithm;
s6, according to the combined loss L ═ Lm+λLaSetting lambda as an auxiliary task weight factor;
s7, calculating a confidence value C (m) of the prediction probability through the multi-classification sub-network;
and S8, judging whether the confidence value C (m) is greater than a set experience threshold value CT, if so, outputting the result as an unknown steganography algorithm, and if not, outputting the type of the steganography algorithm.
Two related steganalysis tasks are constructed in the application: a two-classification task to distinguish between normal audio (Cover) and secret audio (Stego) images and a multi-classification task to distinguish between the types of the secret audio steganography algorithm. Of the two tasks, the two classification tasks for distinguishing the normal audio and the secret audio are the main focus of the work and can be regarded as a main task (maincast).
In particular, the role of the feature extraction sub-network is to adaptively extract steganalysis features from the input audio data. Setting a reasonable preprocessing layer in a CNN (Convolutional Neural Networks) steganalysis model can often improve the steganalysis performance of the network, so we use an audio preprocessing layer based on differential filtering design at the beginning of a feature extraction subnetwork, which is composed of 4 1 × 5 convolution kernels D1-D4, and the initial weights are respectively: d1 ═ 1, -1,0,0, 0], D1 ═ 1, -2,1,0,0], D1 ═ 1, -3, 3, 1, 0], D1 ═ 1, -4,6, -4, 1;
the audio pre-processing layer is followed by 5 concatenated convolution groups, namely, the 1 st convolution group, the 2 nd convolution group, the 3 rd convolution group, the 4 th convolution group and the 5 th convolution group.
The 1 st convolution group includes a 1 × 1 first convolution layer, a 1 × 5 second convolution layer, and a 1 × 1 third convolution layer; the 1 st convolutional layer is activated by using a Truncation Linear Unit (TLU), the TLU can inhibit the activation of an overlarge positive value zone and keep certain activation capability in a negative value zone compared with a linear rectification activation unit ReLU commonly used in a deep learning voice recognition task, and compared with another commonly used TanH activation unit, the TLU has a larger activation interval and keeps certain gradient in the activation interval, so that the risk of gradient disappearance can be reduced when a network is trained. In addition, the other convolution layers of the 1 st convolution group do not carry out activation processing, and the pooling operation is cancelled, so that the aim is to more effectively capture the weak information brought by steganography.
The 2 nd, 3 rd, 4 th and 5 th convolution groups all contain a 1 × 5 convolution layer, a 1 × 1 convolution layer and a mean pooling layer, wherein the last mean pooling layer of the 5 th convolution group module is replaced by a global mean pooling (global average Pooling) layer for the purpose of fusing global features.
The feature extraction sub-network further comprises a feature output layer, the feature output layer is composed of a fully connected layer with 256 neurons, and the feature output layer finally outputs 256-dimensional steganalysis feature vectors F.
The detailed parameters for each subnetwork are shown in the following table:
the numerical meanings in the tables are exemplary: 64 × (1 × 5), ReLU, a 1 × 5 convolution kernel with the parameter setting for the convolutional layer being output channel 64, and activating the output using the ReLU; FC-256 represents a fully connected layer with 256 neurons.
The two classification sub-networks are next to the feature output layer and are composed of full connection containing 128 neuronsAnd (3) layer composition. Feature vector F outputs a binary steganographic predictive probability vector through the subnetwork
Calculating binary steganographic label vector y ═ y after the binary steganographic label vector is coded with One-hot
0,y
1](y
i∈ {0,1}, i represents the class index, and when the value on the class index i is 1, the cross entropy loss L representing the data label class i) is I)
m,
And finally, updating the network parameters through a back propagation error and gradient descent algorithm.
The structure of the multi-classification sub-network is a fully-connected layer of two-layer cascade, and two sides of the cascade respectively comprise 128 neuron structures and 64 neuron structures. The feature vector F outputs the prediction probability value of the steganography algorithm type through the sub network
Calculating the steganographic class label m ═ m after encoding with One-hot
0,m
1,…,m
M-1](representing M different steganographic algorithms of the classes contained in the training set data) cross entropy loss L
a,
And finally, updating the network parameters through a back propagation error and gradient descent algorithm.
The optimization problem required by the whole network is the comprehensive loss L (L) of the main task loss and the auxiliary task lossm+λLaWherein λ is an auxiliary task weight factor, which determines the importance of the auxiliary task to the main task, and the larger λ is, the larger the guidance of the auxiliary task to the main task training is, and the larger interference information brought by the auxiliary task is, so it is also important to set a reasonable λ.
Finally, the multiclassification subnetwork also includes a Softmax activation layer, which outputs a predicted probability p (m) that can be considered for each classification category
k) The more concentrated the predictive probability distribution, the more representative of the network predictionThe larger the credibility is, the information entropy can reflect the concentration degree of the information, so that the confidence value of the prediction probability is calculated and output according to the information entropy
As known from the knowledge of information theory, the confidence value C (M) obtains the maximum value logM (C (M)) when the output probability is uniformly distributed (namely the prediction probability of all algorithm types is 1/M)maxLog M, from which an empirical confidence threshold CT is set, CT 0.5c (M)max。
When the confidence value C (m) is greater than CT, the prediction probability distribution is considered to be uniform, and the network has low prediction confidence for each type of algorithm, so that the algorithm can be considered not to be included in the training data, namely, the unknown steganography algorithm. Conversely, when the confidence value C (m) is less than CT, the prediction probability distribution is more concentrated, and the type with the highest prediction probability is selected as the output type of the data.
While embodiments of the invention have been shown and described, it will be understood by those skilled in the art that: various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.