A method for detecting Android malware based on big data
Through big data analysis and deep learning technology, hyperinformation neural network and convolutional neural network are used to extract and detect features of Android software, solving the accuracy and user experience problems of malware detection in Android system, and achieving efficient and automated Security detection to adapt to new threats and scenario needs.
Patent Information
- Application Number
- CN202410699790.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-31
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-05-31
AI Technical Summary
The existing technology has problems of accuracy and inefficiency in Android systems. Traditional authentication methods are difficult to deal with evolving threats, and the user experience is poor, requiring tedious operational steps.
Using a big data-based method, the sample data of Android software is extracted and analyzed through pre-trained hyperinformation neural network and fully connected network, combined with information decision-making convolutional neural network and improved deep convolutional neural network, and used Compare Loss loss Functions are trained to implement automated malware detection.
Improves the security protection capabilities of Android systems, accurately identify malware, reduces false alarm rates, optimizes user experience, and realizes automatic background detection without additional user operations. It is adaptable and scalable, and can cope with the evolution of malware and new types. attack.
Smart Images

Figure CN118503967B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to big data processing technology and convolutional neural networks, and in particular to an Android malware detection method based on big data. Background Art
[0002] With the rapid development of mobile Internet, the Android system, as one of the most widely used mobile operating systems, has received increasing attention for its security issues. Malware, network attacks and other security threats are emerging in an endless stream, posing a serious threat to users' privacy and property security. Therefore, a big data-based Android malware detection method has emerged, providing strong support for the security protection of the Android system.
[0003] The core idea of this method is to use big data technology to analyze and mine massive amounts of Android application data to discover potential malware. By collecting various information about Android applications, such as application behavior, permission requests, network interactions, etc., a huge data set is constructed. Then, algorithms are used to process and analyze this data to identify the characteristics and behavior patterns of malware.
[0004] During the data analysis process, the present invention pays special attention to the code structure of the application, and the application behavior is used to infer whether there is malware activity. For example, some malware may secretly collect user information, send malicious requests, or perform other improper operations in the background, and these behaviors can be revealed through data analysis.
[0005] Compared with traditional identity authentication methods, the Android malware detection method based on big data has higher accuracy and efficiency. Traditional identity authentication methods often rely on fixed verification rules and patterns, which are difficult to cope with evolving threats. The present invention can adaptively learn and identify new malware features through deep learning and big data analysis, thereby improving the security of the system.
[0006] At the same time, this method can also optimize the user experience. Although the traditional user active authentication method improves security, it increases the user's verification steps and complexity. The Android malware detection method based on big data can automatically perform security detection in the background without the user performing additional operations, thereby improving the convenience and smoothness of the login process. Summary of the invention
[0007] In order to solve the problems existing in the prior art, the present invention provides an Android malware detection method based on big data, comprising the following steps:
[0008] Inputting sample data of the Android software to be detected into the pre-trained hyper-information neural network to obtain first feature data of the sample data;
[0009] Inputting the first feature data into a pre-trained fully connected network to obtain second feature data of the sample data;
[0010] Position encoding is performed on the second feature data to obtain a unique vector representation of each node;
[0011] The unique vector representation of the node is input into the prediction network composed of the information decision convolutional neural network model, the boosted deep convolutional neural network and the global maximum pooling layer to obtain the detection result.
[0012] Furthermore, the hyperinformation neural network is composed of a cascaded multi-layer structure, and the data processing of the l+1 layer includes:
[0013]
[0014] Among them, H (l+1) is the output of the l+1th layer of the hyperinformation neural network, is the adaptive weight vector of the lth layer of the hyperinformation neural network; ⊙ represents matrix element multiplication; b (l) is the learnable bias of the lth layer of the hyperinformation neural network.
[0015] Furthermore, the adaptive weight vector of the lth layer of the hyperinformation neural network The elements in are updated according to back propagation. When initialized, the value of the i-th element in the adaptive weight vector is the accumulation of the adaptive weights between the i-th node and all its neighboring nodes. The adaptive weight between the i-th node and its j-th neighboring node is expressed as:
[0016]
[0017] Among them, sim(x i ,x j ) represents the similarity measure between the i-th node and the j-th node, and N(i) represents the set of neighbor nodes of the i-th node.
[0018] Furthermore, the fully connected network is composed of a plurality of cascaded residual blocks, each residual block includes two cascaded linear fully connected layers, and the output of the second linear fully connected layer is added to the input of the residual as the output of the parameter block.
[0019] Furthermore, the processing of input data by a residual block includes:
[0020] Z1=W1·x+b1
[0021] f1(x)=σ(Z1)
[0022] Z2=W2·f1(x)+b2
[0023] f2(x)=σ(Z2)
[0024] y=f2(x)+x
[0025] Among them, x is the input of the residual layer; W1 is the weight of the first fully connected layer, b1 is the bias of the first fully connected layer, Z1 is the output of the first fully connected layer, f1(x) is the linear transformation output of the first fully connected, and σ is the nonlinear activation function; W2 is the weight of the second fully connected layer, b2 is the bias of the second fully connected layer, Z2 is the output of the second fully connected layer, f2(x) is the linear transformation output of the second fully connected; y is the output of the residual block.
[0026] Furthermore, the Compare Loss loss function is used to propagate the direction of the hyper-information neural network to complete the training of the hyper-information neural network. The Compare Loss loss function is expressed as:
[0027]
[0028] Among them, f(x) is the Compare Loss loss function; α is the relative importance weight factor of similar features, β is the relative importance weight factor of dissimilar features; y is the mark of similarity. When the samples corresponding to sample feature xi and sample feature yi are both positive samples or negative samples, the samples are similar, then y=1, otherwise y=0; D represents the sample set; B is the set threshold.
[0029] Furthermore, the prediction network includes two branches. The first branch is composed of a cascaded information decision convolutional neural network model and a boosted deep convolutional neural network model, and the second branch is composed of a cascaded boosted deep convolutional neural network and an information decision convolutional neural network model. The outputs of the two branches are spliced together and input into the global maximum pooling layer to obtain the final prediction result.
[0030] Furthermore, the process of processing data by the information decision convolutional neural network model includes the following steps:
[0031] Calculate the attention coefficient between the features of the i-th node and the j-th node, that is:
[0032] e ij =LeakyReLU(a T [W 1 h i ||W 2 h j ])
[0033] Among them, a T represents the transpose of the weight vector a of the attention mechanism; h i 、h jare the feature representations of i and j respectively; LeakyReLU(·) is the LeakyReLU activation function; W 1 , W 2 are the first and second learnable weight matrices in the information decision convolutional neural network model respectively; || represents the concatenation operation;
[0034] The softmax function is used to normalize the attention coefficient, that is:
[0035]
[0036] Among them, N i represents the neighbor set of node i;
[0037] The normalized attention coefficient of the current node and the neighboring nodes of the current node are used to update the current node, that is:
[0038]
[0039] Among them, h i ′ is the node i after being processed by the information decision convolutional neural network model; W 3 It is the third learnable weight matrix in the information decision convolutional neural network model.
[0040] Furthermore, the process of improving the deep convolutional neural network to process data includes the following steps:
[0041] f(v)=σ(W1·x v +W2·AGG(v,N(v))+W3·HIST(v)+W4·h emb (v))
[0042]
[0043] Where f(v) represents the node v obtained after the boosted deep convolutional neural network processing; σ(·) is the ReLU activation function; W1, W2, W3, and W4 are the first, second, third, and fourth learnable weight matrices in the boosted deep convolutional neural network, respectively; x v is the feature vector of node v input into the deep convolutional neural network; AGG(v,N(v)) represents the aggregation of all nodes in the neighbor node set N(v) of node v; HIST(v) represents the historical feature vector of node v input into the deep convolutional neural network; h emb (v) is the embedding representation of node v.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] 1. Improved the security protection capability of the Android system: By collecting and analyzing a large amount of Android application data, this method can accurately identify the characteristics and behavior patterns of malware, thereby effectively preventing malware from infringing on user privacy and property security. Compared with traditional security detection methods, this method has higher detection accuracy and lower false alarm rate.
[0046] 2. Optimized user experience: Traditional security detection methods often require users to perform cumbersome operations, such as installing additional security software and making complex settings, which brings inconvenience to users. However, this method automatically analyzes user behavior data in the background to achieve intelligent detection of malware without the need for users to perform additional operations, greatly improving the user experience.
[0047] 3. Adaptability and scalability: With the continuous evolution of malware and the emergence of new attack methods, traditional security detection methods may not be able to respond in time. This method is based on machine learning and big data analysis, and can continuously learn and adapt to new malware features to maintain effective detection of the latest security threats. At the same time, this method can also be expanded and optimized according to actual needs to meet the security detection needs in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 This is a flow chart of an Android malware detection method based on big data of the present invention;
[0049] Figure 2 This is a schematic diagram of the data processing flow during the training process of an Android malware detection method based on big data according to the present invention;
[0050] Figure 3 The present invention is a flowchart of implementing data processing in an Android malware detection method based on big data. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] The present invention provides an Android malware detection method based on big data, comprising the following steps:
[0053] Inputting sample data of the Android software to be detected into the pre-trained hyper-information neural network to obtain first feature data of the sample data;
[0054] Inputting the first feature data into a pre-trained fully connected network to obtain second feature data of the sample data;
[0055] Position encoding is performed on the second feature data to obtain a unique vector representation of each node;
[0056] The unique vector representation of the node is input into the prediction network composed of the information decision convolutional neural network model, the boosted deep convolutional neural network and the global maximum pooling layer to obtain the detection result.
[0057] In the present invention, each software is taken as a node to be detected, the Android application code structure, third-party library and components corresponding to the software are taken as the characteristics of the node, and each sample data is taken as a node. By processing the characteristics of the node, it is predicted whether the node is malware. In the process of training networks such as hyperinformation neural networks, the present invention takes the sample data corresponding to malware in historical data as negative samples, and the sample data corresponding to non-malware as positive samples.
[0058] In the real-time processing stage, the present invention uses a pre-trained hyper-information neural network, a fully connected network, and a prediction network to process the input data to obtain a prediction result.
[0059] The Super Information Neural Network (ISNN) is a multi-layer feedforward neural network, where each layer contains multiple neurons. Each neuron receives input from the neurons in the previous layer and produces output through weighted summation and activation function. The key feature of ISNN is its adaptive weight mechanism, which allows the network to dynamically adjust weights based on the characteristics of the input data.
[0060] The node set in ISNN is V. In the present invention, a software is regarded as a node. The edge set is E. In the present invention, the similarity between nodes is the edge relationship between nodes. The node feature vector is x i , the node at the lth layer is represented by The adaptive weight is σ represents the ReLU activation function, and the bias term is b (l) , each layer of adaptive weights is updated through back propagation, and the adaptive weights are initialized according to the node feature vector. The initialization process includes:
[0061]
[0062] Among them, w ij represents the adaptive weight between the feature vector of the i-th node and the feature vector of the j-th node, sim(x i ,x j ) represents the similarity measure between the feature vector of the i-th node and the feature vector of the j-th node, sim(x i ,xk ) represents the similarity measure between the feature vector of the ith node and the feature vector of the kth node. As an optional embodiment, cosine similarity can be used as the similarity function; N(i) is the neighbor set of the ith node. In this embodiment, the K nodes with the highest similarity to the current node are used as the neighbor set of the current node.
[0063] Forward propagation refers to the output of the lth layer node in ISNN To update, the update process is:
[0064]
[0065] Back propagation and optimization means that during the training process, ISNN uses the back propagation algorithm to optimize the network weights and bias items. First, the loss function value between the network output and the true label is calculated. Then, the gradient of the loss function to each layer of weights and bias items is calculated by the chain rule. Finally, the gradient descent algorithm is used to update the weights and bias items to minimize the loss function.
[0066] The above update is the update process of a certain node. The update of all input nodes is expressed in matrix form, which is expressed as:
[0067]
[0068] Among them, H (l+1) is the output of the l+1th layer of the hyperinformation neural network, and all nodes Composition H (l+1) ; is the adaptive weight vector of the lth layer of the hyperinformation neural network, all constitute ⊙ represents matrix element multiplication; W (l) is the learnable matrix of the lth layer of the hyperinformation neural network; b (l) is the learnable bias of the lth layer of the hyperinformation neural network, and the Composition b (l) .
[0069] The ReLU activation function used in this embodiment is an activation function commonly used in artificial neural networks, which refers to the ramp function in mathematics, that is:
[0070] f(x)=max(0,x)
[0071] The above formula means that when the input x is greater than zero, the output is equal to the input x, and when the input is less than zero, the output is zero. In simple terms, it returns all negative values to zero, while positive values remain unchanged.
[0072] The ReLU activation function enables neural networks to learn very complex functional relationships and is very efficient in calculation. It can help neural networks better handle various types of data and help avoid some gradient disappearance problems.
[0073] The feature results output by the hyperinformation neural network are passed to the fully connected network, and the Compare Loss training is started to obtain a more accurate feature vector. The Compare Loss loss function is expressed as:
[0074]
[0075] Among them, f(x) is the Compare Loss loss function; α is the relative importance weight factor of similar pair features, β is the relative importance weight factor of dissimilar pair features; y is the mark of similarity. When the samples corresponding to sample feature xi and sample feature yi are both positive samples or negative samples, the samples are similar, and y=1 at this time, otherwise y=0; D represents the sample set; B is the set threshold, which is used to judge that the similarity of similar pair features must be greater than the set threshold before the loss function of similar pair features takes effect. At the same time, y=1 is established only when the similarity of the two features is greater than the threshold. Represents sample feature xi and sample feature yi. In this embodiment, Euclidean distance can be used for calculation. Those skilled in the art can change the calculation method of the similarity according to the adaptability of the selected features.
[0076] The feature results output by the super information neural network (ISNN) are passed to the fully connected network, and the details of the fully connected network with residual connection are as follows:
[0077] Define residual blocks. Residual blocks allow input features to skip one or more layers and pass to deeper networks. Each residual block consists of two fully connected layers and a cross-layer residual connection.
[0078] The residual block includes two fully connected layers. The weight of the first fully connected layer is W1 and the bias is b1. The weight of the second fully connected layer is W2 and the bias is b2. Both fully connected layers use nonlinear activation functions.
[0079] The output of the first fully connected layer can be expressed as:
[0080] Z1=W1·x+b1
[0081] f1(x)=σ(Z1)
[0082] The input of the second fully connected layer is the output of the first fully connected layer (f1(x)), and its output can be expressed as:
[0083] Z2=W2·f1(x)+b2
[0084] f2(x)=σ(Z2)
[0085] The residual connection ensures that the dimensions of (x) and (f2(x)) are consistent through linear transformation. In this embodiment, they can be directly added to obtain the output of the residual block:
[0086] y=f2(x)+x
[0087] y is the final output of the residual block, which fuses the features passed through the fully connected layer and the features directly passed through the residual connection.
[0088] In order to build a deeper network, multiple residual blocks are stacked together, and the output of each residual block will be used as the input of the next residual block. By passing and fusing feature information layer by layer, the network can learn higher-level feature representations.
[0089] Using residual connections can help overcome some difficulties in deep neural network training, improve the training efficiency and performance of the network, and make deeper network structures easier to train and optimize. In traditional deep neural networks, the gradient may gradually become smaller during the back propagation process, causing the gradient to disappear when passing through layers. Using residual connections, the gradient can be propagated more easily because the original input is directly added to the output of the subsequent layer. Even if the weight of the subsequent layer is zero, the output of the residual block will still be consistent with the original input, thus avoiding the problem of gradient disappearance. At the same time, it can accelerate training convergence. Since the gradient is easier to propagate, using residual connections can accelerate the training convergence of the network. This means that less training time and less data are needed to achieve the same performance level. The most critical thing is that it can improve network performance. Residual connections allow the network to learn residual functions instead of original functions. This means that the network can focus more on learning the remaining part, thereby improving the performance and representation ability of the network. By allowing the network to skip one or more layers to pass features, residual connections can help the network better learn the representation of data, thereby improving the generalization ability of the network.
[0090] The feature vector obtained by residual block processing is embedded using position encoding. According to the category information of the node, a unique vector is generated for each node. The position encoding embedding method is:
[0091] PE(pos,2i)=sin(pos / 10000 (2i / d model) )
[0092] PE(pos,2i+1)=cos(pos / 10000 (2i / d model) )
[0093] Among them, PE represents positional encoding, PE(pos,2i) represents the position encoding of the position pos of the current character in the input sequence on the i-th dimension of the embedding vector; d model represents the dimension of the embedding vector.
[0094] For each dimension of the embedding vector (indexed by i), a sine value (for even-indexed dimensions) and a cosine value (for odd-indexed dimensions) are calculated. These two values are combined to form a unique vector for each position. The periodicity of the sine and cosine functions allows the model to capture relative and absolute relationships between different positions.
[0095] After generating the position encoding, we need to combine it with the node's category information (i.e., the node's feature vector). This can be done by concatenating the position encoding with the node's feature vector. In this way, each node will have a unique vector representation that contains both category information and position information.
[0096] These encoded node vectors are fed into the IDCN model for training and optimization. During the training process, the model will learn how to use these vectors to distinguish nodes of different categories and optimize its performance. Through continuous iteration and optimization, a model that can accurately identify node categories can be obtained.
[0097] In this embodiment, the prediction network includes two branches. The first branch is composed of a cascaded information decision convolutional neural network model and a boosted deep convolutional neural network model, and the second branch is composed of a cascaded boosted deep convolutional neural network and an information decision convolutional neural network model. The outputs of the two branches are spliced together and input into the global maximum pooling layer to obtain the final prediction result.
[0098] The core idea of the Information Decision Convolutional Neural Network (IDCN) model is to combine the graph attention mechanism with the message propagation process of GCN, so that the model can dynamically learn the importance weights between nodes, so as to better focus on key graph structure information.
[0099] The graph attention mechanism allows each node to dynamically calculate its attention weight based on the characteristics of neighboring nodes. In this way, the model can more flexibly capture the dependencies between nodes and ignore those neighboring nodes that have little impact on the target node. By further aggregating the weighted message propagation results, enhanced node representation can be obtained to more accurately represent the structural information of the graph. The process of processing the input by the Information Decision Convolutional Neural Network (IDCN) model includes the following steps:
[0100] For any node i in the graph, the set of its neighbor nodes is denoted by N iFor each neighbor node j∈N i , calculate an attention coefficient e ij , the attention coefficient reflects the importance of node j to node i. This coefficient is usually calculated based on node features and graph structure. The calculation formula of the attention coefficient is as follows:
[0101] e ij =LeakyReLU(a T [W 1 h i ||W 2 h j ])
[0102] Among them, a T represents the transpose of the weight vector a of the attention mechanism; h i 、h j are the feature representations of i and j respectively; LeakyReLU(·) is the LeakyReLU activation function; W 1 , W 2 are the first and second learnable weight matrices in the information decision convolutional neural network model respectively; || represents the concatenation operation;
[0103] The softmax function is used to normalize the attention coefficient, that is:
[0104]
[0105] After obtaining the attention coefficients, the features of neighboring nodes are weighted and aggregated according to these coefficients to update the feature representation of node i:
[0106]
[0107] Among them, h i ′ is the node i after being processed by the information decision convolutional neural network model; W 3 It is the third learnable weight matrix in the information decision convolutional neural network model;
[0108] Through the above process, the new feature representation h of node i i ′ not only considers its own features, but also the features of its neighboring nodes, and weights them according to the importance of the neighboring nodes.
[0109] In the Boosted Deep Convolutional Neural Network (PBCN) model, a more advanced aggregation mechanism is introduced to capture the personalized features or dynamic changes of nodes in the graph. This mechanism relies on the historical information of the node, the features of the neighboring nodes, or the structure of the entire graph. The formula of the aggregation function is as follows:
[0110] f(v)=σ(W1·x v+W2·AGG(v,N(v))+W3·HIST(v)+W4·h emb (v))
[0111]
[0112] Where f(v) represents the node v obtained after the boosted deep convolutional neural network processing; σ(·) is the ReLU activation function; W1, W2, W3, and W4 are the first, second, third, and fourth learnable weight matrices in the boosted deep convolutional neural network, respectively; x v is the node v of the input deep convolutional neural network; f(u) represents the node u obtained after being processed by the boosted deep convolutional neural network. When the node has not been processed by the boosted deep convolutional neural network, it is replaced by the initial features of the node u, that is, the features of the node u input into the boosted deep convolutional neural network.
[0113] In the aggregate function, W1·x v This part converts the current state or feature vector x of node v into v A linear transformation is performed through a learnable weight matrix W1, which mainly considers the current state of the node; W2 · AGG (v, N (v)) first uses the aggregation function AGG to aggregate the feature vectors x of all neighbor nodes u of node v u Aggregated into a single vector, this aggregated vector is then linearly transformed through a learnable weight matrix W2, mainly considering the state of the neighboring nodes of the main node; W3·HIST(v) This part linearly transforms the historical information vector of node v through a learnable weight matrix W3, considering the node history information; W4·h emb (v) The structural information vector of graph G is linearly transformed through a learnable weight matrix W4, taking into account the information of the nodes in the graph structure.
[0114] In this embodiment, in addition to the aggregation mechanism, PBCN also innovates the node update mechanism, including introducing more complex transformations and using the attention mechanism to dynamically adjust the weight of the node. The formula for updating the node at the k+1th iteration is as follows:
[0115]
[0116] Among them, UPDATE (k) is a learnable update function that generates a new embedding representation based on the current embedding of the node and the aggregated features. And the neighbor feature after the k-th iteration aggregation AGGREGATE PBCN (k)(v) The merged feature vector is obtained by concatenation, and the merged feature vector is input into the fully connected layer. After linear transformation and RELU activation function processing, the embedding vector of the k+1th iteration is obtained.
[0117] In order to prevent overfitting and improve the generalization ability of the model, PBCN adopts some innovative regularization strategies and adds regularization terms to the loss function as shown below:
[0118] L PBNN =L original +λ1L L1 +λ2L L2 +λ3L graph_reg
[0119] Among them, L original is the original loss function of PBCN, and the existing loss functions such as cross entropy can be used; L L1 , L L2 is the regularization term, L graph_reg is a regularization term based on the graph structure, and λ1, λ2, and λ3 are regularization parameters.
[0120] In the prediction network, the outputs of the two branches are concatenated together as the input of the global maximum pooling layer, and the global maximum pooling operation outputs the detection result. Figure 2 In the training process of this embodiment, the negative samples are input into the second feature data extraction network formed by the cascade of the first ISNN and the first FC net. The extracted negative samples are spliced with the positive samples as the second feature data extraction network formed by the cascade of the second ISNN and the second FC net. The extracted second feature data is positionally encoded, and the feature data and the position encoding are spliced as the input of the prediction network. The prediction network of this embodiment includes two branches. The outputs of the two branches are spliced and global maximum pooling is performed to obtain the prediction result. The two branches are constrained by the idea of regularization during the training process.
[0121] like Figure 3 When processing the data to be detected, the present invention uses a pre-trained ISNN and a fully connected layer network, that is, a second ISNN and a second FC net to process and then perform position encoding, and then splices the position encoding with the second feature data output by the fully connected layer network as the input of the prediction network, and the prediction network outputs predicted data.
[0122] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A big data-based Android malware detection method, It is characterized in that The specific steps include: Inputting sample data of the Android software to be detected into the pre-trained hyper-information neural network to obtain first feature data of the sample data; The hyperinformation neural network consists of a cascaded multi-layer structure. The data processing of the l+1 layer includes: Among them, H (l+1) is the output of the l+1th layer of the hyperinformation neural network, is the adaptive weight vector of the lth layer of the hyperinformation neural network; ⊙ represents matrix element multiplication; b (l) is the learnable bias of the lth layer of the hyperinformation neural network, σ is the nonlinear activation function; Adaptive weight vector of the lth layer of the hyperinformation neural network The elements in are updated according to back propagation. When initialized, the value of the i-th element in the adaptive weight vector is the accumulation of the adaptive weights between the i-th node and all its neighboring nodes. The adaptive weight between the i-th node and its j-th neighboring node is expressed as: Among them, sim(x i ,x j ) represents the similarity measure between the i-th node and the j-th node, and N(i) represents the set of neighbor nodes of the i-th node; The first feature data is input into the pre-trained fully connected network to obtain the second feature data of the sample data; the fully connected network is composed of a plurality of cascaded residual blocks, each residual block includes two cascaded linear fully connected layers, and the output of the second linear fully connected layer is added to the input of the residual as the output of the parameter block; the processing process of a residual block on the input data includes: Z 1 =W 1 ·x+b 1 f 1 (x)=σ(Z 1 ) Z 2 =W 2 ·f 1 (x)+b 2 f 2 (x)=σ(Z 2 ) y=f 2 (x)+x Among them, x is the input of the residual layer; W 1 is the weight of the first fully connected layer, b 1 is the bias of the first fully connected layer, Z 1 is the output of the first fully connected layer, f 1 (x) is the first fully connected linear transformation output; W 2 is the weight of the second fully connected layer, b 2 is the bias of the second fully connected layer, Z 2 is the output of the second fully connected layer, f 2 (x) is the second fully connected linear transformation output; y is the output of the residual block; The Compare Loss loss function is used to propagate the direction of the hyper-information neural network and complete the training of the hyper-information neural network. The Compare Loss loss function is expressed as: Among them, f(x) is the Compare Loss function; α is the relative importance weight factor of similar features, β is the relative importance weight factor of dissimilar features; y is the mark of similarity, when the sample feature x i With sample feature y i If the corresponding samples are all positive samples or negative samples, the samples are similar, and y=1, otherwise y=0; D represents the sample set; B is the set threshold; the second feature data is position-encoded to obtain a unique vector representation of each node; The unique vector representation of the node is input into the prediction network composed of the information decision convolutional neural network model, the boosted deep convolutional neural network and the global maximum pooling layer to obtain the detection result, that is, the prediction network includes two branches, the first branch is composed of the cascaded information decision convolutional neural network model and the boosted deep convolutional neural network, and the second branch is composed of the cascaded boosted deep convolutional neural network and the information decision convolutional neural network model. The outputs of the two branches are spliced together and input into the global maximum pooling layer to obtain the final prediction result; The process of improving the deep convolutional neural network to process data includes the following steps: f(v)=σ 1 (W 1 ·x v +W 2 ·AGG(v,N(v))+W 3 ·HIST(v)+W 4 ·h emb (v)) Where f(v) represents the node v obtained after processing by the boosted deep convolutional neural network; σ 1 (·) is the ReLU activation function; W 1 , W 2 , W 3 , W 4 They are respectively used to improve the first, second, third, and fourth learnable weight matrices in the deep convolutional neural network; x v is the feature vector of node v input into the deep convolutional neural network; AGG(v,N(v)) represents the aggregation of all nodes in the neighbor node set N(v) of node v; HIST(v) represents the historical feature vector of node v input into the deep convolutional neural network; h emb (v) is the embedding representation of node v.
2. According to the big data-based Android malware detection method of claim 1, It is characterized in that The process of processing data by the information decision convolutional neural network model includes the following steps: Calculate the attention coefficient between the features of the i-th node and the j-th node, that is: and ij =LeakyReLU(a T [IN 1 h i ||In 2 h j ]) Among them, a T represents the transpose of the weight vector a of the attention mechanism; h i 、h j are the feature representations of i and j respectively; LeakyReLU(·) is the LeakyReLU activation function; W 1 , W 2 are the first and second learnable weight matrices in the information decision convolutional neural network model respectively; || represents the concatenation operation; The softmax function is used to normalize the attention coefficient, that is: Among them, N i represents the neighbor set of node i; The normalized attention coefficient of the current node and the neighboring nodes of the current node are used to update the current node, that is: Among them, h′ i is the node i after being processed by the information decision convolutional neural network model; W 3 It is the third learnable weight matrix in the information decision convolutional neural network model.
