Physical entropy source security assessment method for multi-source information fusion

Abstract

The invention belongs to the technical field of information security and random number generators, and aims to solve the problem that the traditional minimum entropy estimation method does not consider the influence of multi-source data such as environmental factors on the random number quality of a physical entropy source. The invention provides a multi-source information fusion physical entropy source security assessment method, which comprises the following steps of: acquiring and processing physical entropy source data to obtain a multi-source fusion data set; performing feature extraction on the multi-source fusion data set through a variational auto-encoder network, and outputting potential representation as data after feature extraction; inputting the potential representations divided into the test sets into the verified hybrid deep learning neural network model to obtain a prediction result of the physical entropy source; calculating a global prediction probability and a local prediction probability based on the prediction result; obtaining a minimum entropy evaluation result based on the global prediction probability and the local prediction probability; and judging the safety of the physical entropy source based on a minimum entropy evaluation result. The problem that existing physical entropy source evaluation is low in accuracy can be solved.

Description

Technical Field

[0001] The invention belongs to the technical field of information security and random number generators, and in particular relates to a physical entropy source security assessment method for multi-source information fusion. Background Art

[0002] Random number generators are widely used in spread spectrum communications, scientific computing, numerical analysis, identity authentication, etc., especially in cryptographic applications and cryptographic systems. They can be divided into two types: true random number generators and pseudo-random number generators. Pseudo-random number generators use algorithms to generate digital sequences with approximate randomness and potential predictability. Unlike pseudo-random number generators, true random number generators generate true random number sequences based on unpredictable physical random phenomena or processes (i.e., physical entropy sources), so true random numbers have better security. Therefore, a high-accuracy minimum entropy estimation method is very important for evaluating the quality of random numbers.

[0003] However, traditional minimum entropy estimation methods based on statistics and deep learning only consider univariate data (i.e., the time series of random numbers), without considering the impact of multi-source data such as environmental factors (such as temperature and voltage fluctuations) on the quality of random numbers. Physical entropy sources (such as thermal noise and electronic noise) are very sensitive to changes in temperature, power supply voltage, etc., and their changes will greatly affect the quality of random numbers. This limitation reduces the accuracy of minimum entropy estimation. Summary of the invention

[0004] In order to solve at least one of the above-mentioned technical problems existing in the prior art, the present invention provides a physical entropy source security assessment method based on multi-source information fusion.

[0005] The present invention is implemented by the following technical solution: a physical entropy source security assessment method based on multi-source information fusion, comprising the following steps:

[0006] The physical entropy source data is obtained and processed to obtain a multi-source fusion data set; the physical entropy source data includes the external temperature, power supply voltage and original data of the output random number sequence of the physical entropy source; the multi-source fusion data set is feature extracted by a variational autoencoder network, a potential representation is output as the data after feature extraction, and the potential representation is divided into a training set, a verification set and a test set; an initial hybrid deep learning neural network model is built, and the model is trained and verified based on the potential representation divided into the training set and the verification set; the potential representation divided into the test set is input into the verified hybrid deep learning neural network model to obtain the prediction result of the physical entropy source; the global prediction probability and the local prediction probability are calculated based on the prediction result; the minimum entropy evaluation result is obtained based on the global prediction probability and the local prediction probability; the safety of the physical entropy source is judged based on the minimum entropy evaluation result.

[0007] Preferably, the variational autoencoder network includes an encoder and a decoder, the variational autoencoder network is trained using an LSTM network, and the loss function is a linear combination of cross entropy loss and KL divergence; the potential representation of the variational autoencoder network output includes the external temperature, power supply voltage and output random number sequence of the physical entropy source after feature extraction.

[0008] Preferably, every 10 consecutive adjacent numbers in the potential representation output by the variational autoencoder network are taken as a group of sequences, the first group of two adjacent groups of sequences is taken as the input sequence of the hybrid deep learning neural network model, and the second group is taken as the output prediction of the first group in the hybrid deep learning neural network model, and the hybrid deep learning neural network model is trained and verified.

[0009] Preferably, the hybrid deep learning neural network model includes a CNN network, an LSTM network, a TPA attention mechanism layer and a fully connected network;

[0010] The CNN network includes a CNN layer, and the input sequence enters the CNN layer for spatial scale feature extraction; the LSTM network includes an LSTM layer, and the LSTM layer outputs the hidden state corresponding to each time step in the input sequence; the TPA attention mechanism layer includes a CNN layer and an evaluation score function, which is used for time series analysis of long sequences; the fully connected network obtains the corresponding output probability through the Softmax activation function;

[0011] The output of the CNN layer of the CNN network is connected to the input of the LSTM layer of the LSTM network, the output of the LSTM layer of the LSTM network is connected to the input of the TPA attention mechanism layer, and the output of the TPA attention mechanism layer is connected to the input of the fully connected network.

[0012] Preferably, in the LSTM layer, the time series fed by the CNN network is received, and the hidden state corresponding to each time step in the input sequence is output, and then all the hidden states are segmented, and all the information is divided into the hidden state H and the current state h t ;

[0013] The hidden state information is input to the TPA temporal attention mechanism layer to extract the long-term pattern in the input information. The row vector of the hidden state H is filtered by the CNN filter. The one-dimensional convolutional CNN layer in the TPA temporal attention mechanism layer is used to perform convolution operations on each row vector. The Relu activation function is selected and H is output. C As the feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer is used as a scoring function to determine the weights of each temporal pattern output by the one-dimensional convolutional CNN layer. The attention is calculated according to the size of the weight. The current LSTM output state h tRepeatedly input to the Multiply layer through the RepeatVector layer to calculate its H C The product of each row of the matrix is ​​summed on the time index axis. After summing, the result is normalized using the Sigmoid activation function to calculate an attention weight a corresponding to each row, which represents H C h to be predicted for each row of the matrix t The intensity of the effect; according to each time series h t The influence of each row is weighted and summed to get the comprehensive time attention v t , v t With the current state h t After merging through the Add layer, we get the output h′ of the attention mechanism layer t .

[0014] Preferably, in the fully connected layer of the fully connected network, a Softmax activation function is used to map the linearly transformed output vector to a set of probability distributions, and the maximum probability therein is selected as the output.

[0015] Preferably, the minimum entropy evaluation result obtained based on the global prediction probability and the local prediction probability is:

[0016] H min =-log 2 (max(P global ,P local ))

[0017] In the formula, H min is the minimum entropy evaluation result, max(P global ,P local ) is the maximum value of the global prediction probability and the local prediction probability, P global is the global prediction probability, P local is the local prediction probability.

[0018] Preferably, the formula for judging the safety of the physical entropy source based on the minimum entropy evaluation result is:

[0019] H min >0.99×M

[0020] In the formula, M=1,2,3….

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The physical entropy source security assessment of multi-source information fusion provided by the present invention can improve the problem of low accuracy of existing physical entropy source assessment. If the minimum entropy estimation result meets the specified size, it is considered that the random numbers generated by the entropy source are unpredictable, thereby ensuring the security of information; the accuracy performance of the minimum entropy assessment of the present invention is superior to the existing statistical and deep learning-based assessment methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 This is a flow chart of minimum entropy estimation of physical entropy source of the present invention;

[0025] Figure 2 It is a variational autoencoder network structure diagram of the present invention;

[0026] Figure 3 This is the structural diagram of the hybrid deep learning model (CNN-TPA-LSTM) of the present invention. DETAILED DESCRIPTION

[0027] In conjunction with the drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. 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 implementation methods obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention. It should be noted that in this specification, relational terms such as first and second are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0029] The present invention provides an embodiment:

[0030] like Figure 1As shown, a method for assessing the safety of a physical entropy source by multi-source information fusion includes the following steps: obtaining and processing physical entropy source data to obtain a multi-source fusion data set; the physical entropy source data includes the external temperature, power supply voltage and original data of the output random number sequence of the physical entropy source; extracting features of the multi-source fusion data set through a variational autoencoder network, outputting a potential representation as data after feature extraction, and dividing the potential representation into a training set, a verification set and a test set; building an initial hybrid deep learning neural network model, training and verifying the model based on the potential representation divided into a training set and a verification set; inputting the potential representation divided into a test set into the verified hybrid deep learning neural network model to obtain a prediction result of the physical entropy source; calculating a global prediction probability and a local prediction probability based on the prediction result; obtaining a minimum entropy evaluation result based on the global prediction probability and the local prediction probability; and judging the safety of the physical entropy source based on the minimum entropy evaluation result.

[0031] In this embodiment, the step of processing the physical entropy source data includes: normalizing the external temperature and power supply voltage data of the physical entropy source, standardizing the random number sequence in the form of an M-bit integer, aligning the external temperature, power supply voltage and random number sequence to make them correspond one-to-one to form a new multi-source fusion data set, using 40% of the entire data set as a training set for building a model, 10% as a validation set for evaluating the model but not updating the model, and 50% as a test set.

[0032] Then it is used as the input to the feature extraction variational autoencoder model as a whole. The potential dimension is set to 10, the intermediate layer dimension is set to 32, the optimizer Adam learning rate is manually set to 0.001, and the loss function is a linear combination of cross entropy loss and KL divergence. Then it is trained through the encoder (LSTM network) and decoder (LSTM network), and the gradient update model is calculated until the loss no longer decreases, and the potential representation z is output.

[0033] like Figure 2 As shown, the variational autoencoder network includes an encoder and a decoder. The variational autoencoder network is trained using an LSTM network, and the loss function is a linear combination of the cross entropy loss and the KL divergence. The potential representation of the variational autoencoder network output includes the external temperature, the supply voltage, and the output random number sequence of the physical entropy source after feature extraction.

[0034] The hybrid deep learning neural network model includes CNN network, LSTM network, TPA attention mechanism layer and fully connected network;

[0035] The CNN network includes a CNN layer, and the input sequence enters the CNN layer for spatial scale feature extraction; the LSTM network includes an LSTM layer, and the LSTM layer outputs the hidden state corresponding to each time step in the input sequence; the TPA attention mechanism layer includes a CNN layer and an evaluation score function, which is used for time series analysis of long sequences; the fully connected network obtains the corresponding output probability through the Softmax activation function;

[0036] The output of the CNN layer of the CNN network is connected to the input of the LSTM layer of the LSTM network, the output of the LSTM layer of the LSTM network is connected to the input of the TPA attention mechanism layer, and the output of the TPA attention mechanism layer is connected to the input of the fully connected network.

[0037] Every 10 consecutive adjacent numbers in the potential representation of the variational autoencoder network output are taken as a set of sequences. The first set of the two adjacent sets of sequences is taken as the input sequence of the hybrid deep learning neural network model, and the second set is taken as the output prediction (label) of the first set in the hybrid deep learning neural network model. Next, the sequence is shifted backward by one position and updated as another set of inputs. Similarly, the next number after the new input sequence is also used as the new output. The shifting process continues until all input sequences and corresponding outputs are generated, and the hybrid deep learning neural network model is trained and verified.

[0038] First, the input sequence enters the CNN layer of the CNN network for further spatial scale feature extraction to make up for the shortcomings of the temporal attention mechanism, and then is sequentially fed to the LSTM layer with an output size of 256, which can output the hidden state corresponding to each time step in the input sequence. After that, the output of the LSTM layer is connected to the TPA attention mechanism layer. The TPA attention mechanism layer with a size of 256 analyzes the information of all previous time steps and selects relevant information to help generate the output. Finally, the output of the temporal attention mechanism enters the fully connected (FC) layer. Since the random number is multi-bit, the FC layer uses the Softmax activation function. The output size of the FC layer is configured to 2 M , which is the number of all possible M-digit numbers when predicting the next value. The deep learning neural network model constructed is as follows Figure 3 shown.

[0039] Among them, in the CNN layer of the CNN network, local features are further extracted from the potential representation input, the number of filters k = 256, the convolution kernel size is 9, and Re lu is used as the activation function.

[0040] In the LSTM layer, the time series fed by the above CNN layer is received. The LSTM has 256 units and outputs the hidden state corresponding to each time step in the input sequence. Then all the hidden states are segmented. All the information is divided into hidden states H = (h 1 ,h 2 , ..., h t-1 ) and the current state h t .

[0041] All hidden state information is input to the TPA temporal attention mechanism layer to extract long-term temporal patterns in the input information. The prediction performance of the model is improved by using CNN filters on the row vectors of H. The one-dimensional convolutional CNN layer in the TPA temporal attention mechanism layer is used to perform convolution operations on each row vector. The one-dimensional convolutional CNN layer has 256 one-dimensional convolution kernels, each with a length of 10. The Relu activation function is selected and the output is H. C As the feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer is used as a scoring function to determine the weights of each temporal pattern output by the one-dimensional convolutional CNN layer. Attention is calculated based on the size of the weights. The current LSTM output state h t Repeatedly input to the Multiply layer through the RepeatVector layer to calculate its H C The product of each row of the matrix is ​​summed on the time index axis. After summing, the result is normalized using the Sigmoid activation function to calculate an attention weight a corresponding to each row, which represents H C h to be predicted for each row of the matrix t According to each time series h t The influence of each row is weighted and summed to get the comprehensive time attention v t , v t With the current state h t After merging through the Add layer, we get the output h′ of the attention mechanism layer t .

[0042] In the fully connected layer of the fully connected network, the Softmax activation function is used to map the linearly transformed output vector to a set of probability distributions. Assume that the linear output of the output layer is z = (z 0 ,z 1 ,…z s ), where s is the size of the sample space, which is equal to the output size of the fully connected layer. The Softmax activation function is defined as Get, where y i Represents the probability corresponding to the i-th output category, satisfying that the sum of all output probabilities is equal to 1, that is In the formula, i=1,2,3….

[0043] In addition, a manually set optimizer Adam learning rate of 0.0005 was introduced to minimize the objective function during the training phase. During training and validation, a batch of 256 samples was input into the model and trained under the loss function cross entropy mechanism. The maximum number of training epochs was set to 1000, and the early stopping monitoring mechanism was introduced. When the loss does not decrease for 7 consecutive epochs, the training is stopped. The validation set loss and the corresponding training weights are recorded during the training phase, and the training weights with the minimum validation error are saved to evaluate the final training weights of the corresponding test set.

[0044] After the model training is completed, the test data set is input, and then the global predictable probability and local predictable probability are calculated by following the same method used in 90B for minimum entropy estimation.

[0045] In a discrete, independent random variable X, the set of possible values ​​is A = {x 1 ,x 2 ,…,x k}, set A is called the sample space or value space of X. It contains all possible values ​​of X (i.e. sample points). A discrete random variable X is a random variable that can take values ​​from each element in set A, and each value has a corresponding probability. i ∈A has, P r (X=x i )=p i , i = 1, ..., k, where p i Corresponding to the y above i , the minimum entropy is defined as:

[0046] The deep learning model predicts the next sample in the sequence based on specific statistical properties of previous samples and provides evaluation results based on the predicted results. The trained model is tested on an untrained dataset, the global and local probabilities of successful predictions are estimated at a 99% confidence level, and the minimum entropy is calculated from the predictions.

[0047] (1) Global prediction probability: Assume that N is the sample size, N T is the number of correct predictions, and r is the number of consecutive correct predictions. r Given by the formula:

[0048] At the 99% confidence level, the global prediction probability P global Calculated as:

[0049]

[0050] (2) Local prediction probability: The local prediction probability is iteratively calculated by the following formula:

[0051]

[0052] Where q = 1-P local , x=x 10 , j=1,2,…,10.

[0053] The minimum entropy evaluation result obtained based on the global prediction probability and the local prediction probability is:

[0054] H min =-log 2 (max(P global ,P local ))

[0055] In the formula, H min is the minimum entropy evaluation result, max(P global ,P local ) is the maximum value of the global prediction probability and the local prediction probability, P global is the global prediction probability, P local is the local prediction probability.

[0056] The formula for judging the safety of the physical entropy source based on the minimum entropy evaluation result is:

[0057] H min >0.99×M

[0058] In the formula, M=1,2,3….

[0059] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A physical entropy source security assessment method based on multi-source information fusion, characterized in that: The following steps are involved: Acquire and process physical entropy source data to obtain a multi-source fusion data set; the physical entropy source data includes the external temperature, power supply voltage and raw data of the output random number sequence of the physical entropy source; Extracting features from the multi-source fusion data set through a variational autoencoder network, outputting a potential representation as feature-extracted data, and dividing the potential representation into a training set, a validation set, and a test set; Build an initial hybrid deep learning neural network model, train and verify the model based on the potential representation divided into training set and validation set; input the potential representation divided into test set into the verified hybrid deep learning neural network model to obtain the prediction result of physical entropy source; Calculate global prediction probability and local prediction probability based on the prediction results; Obtain minimum entropy evaluation results based on global prediction probability and local prediction probability; The safety of the physical entropy source is judged based on the minimum entropy evaluation results.

2. According to claim 1, a physical entropy source security assessment method based on multi-source information fusion is characterized by: The variational autoencoder network includes an encoder and a decoder. The variational autoencoder network is trained using an LSTM network, and the loss function is a linear combination of cross entropy loss and KL divergence. The potential representation of the variational autoencoder network output includes the external temperature, power supply voltage and output random number sequence of the physical entropy source after feature extraction.

3. According to claim 2, a physical entropy source security assessment method based on multi-source information fusion is characterized by: Every 10 consecutive adjacent numbers in the potential representation of the variational autoencoder network output are taken as a sequence. The first group of two adjacent groups of sequences is taken as the input sequence of the hybrid deep learning neural network model, and the second group is taken as the output prediction of the first group in the hybrid deep learning neural network model. The hybrid deep learning neural network model is trained and verified.

4. According to claim 3, a physical entropy source security assessment method based on multi-source information fusion is characterized by: The hybrid deep learning neural network model includes CNN network, LSTM network, TPA attention mechanism layer and fully connected network; The CNN network includes a CNN layer, and the input sequence enters the CNN layer for spatial scale feature extraction; the LSTM network includes an LSTM layer, and the LSTM layer outputs the hidden state corresponding to each time step in the input sequence; the TPA attention mechanism layer includes a CNN layer and an evaluation score function, which is used for time series analysis of long sequences; the fully connected network obtains the corresponding output probability through the Softmax activation function; The output of the CNN layer of the CNN network is connected to the input of the LSTM layer of the LSTM network, the output of the LSTM layer of the LSTM network is connected to the input of the TPA attention mechanism layer, and the output of the TPA attention mechanism layer is connected to the input of the fully connected network.

5. According to claim 4, a physical entropy source security assessment method based on multi-source information fusion is characterized by: In the LSTM layer, the time series fed by the CNN network is received, and the hidden state corresponding to each time step in the input sequence is output. Then all the hidden states are segmented, and all the information is divided into the hidden state H and the current state h t ; The hidden state information is input to the TPA temporal attention mechanism layer to extract the long-term pattern in the input information. The row vector of the hidden state H is filtered by the CNN filter. The one-dimensional convolutional CNN layer in the TPA temporal attention mechanism layer is used to perform convolution operations on each row vector. The Relu activation function is selected and H is output. C As the feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer is used as a scoring function to determine the weights of each temporal pattern output by the one-dimensional convolutional CNN layer. The attention is calculated according to the size of the weight. The current LSTM output state h t Repeatedly input to the Multiply layer through the RepeatVector layer to calculate its H C The product of each row of the matrix is ​​summed on the time index axis. After summing, the result is normalized using the Sigmoid activation function to calculate an attention weight a corresponding to each row, which represents H C h to be predicted for each row of the matrix t The intensity of the effect; according to each time series h t The influence of each row is weighted and summed to get the comprehensive time attention v t , v t With the current state h t After merging through the Add layer, we get the output h′ of the attention mechanism layer t .

6. According to claim 4, a physical entropy source security assessment method based on multi-source information fusion is characterized by: In the fully connected layer of the fully connected network, the Softmax activation function is used to map the linearly transformed output vector to a set of probability distributions, and the maximum probability is selected as the output.

7. According to claim 4, a physical entropy source security assessment method based on multi-source information fusion is characterized by: The minimum entropy evaluation result obtained based on the global prediction probability and the local prediction probability is: H min =-log2(max(P global ,P local )) In the formula, H min is the minimum entropy evaluation result, max(P global ,P local ) is the maximum value of the global prediction probability and the local prediction probability, P global is the global prediction probability, P local is the local prediction probability.

8. The method for physical entropy source security assessment based on multi-source information fusion according to claim 7 is characterized by: The formula for judging the safety of the physical entropy source based on the minimum entropy evaluation result is: H min >0.99×M In the formula, M=1,2,3….

Images