A multi-source information fusion physical entropy source security evaluation method

Abstract

The present application belongs to the technical field of information security and random number generator, and aims to solve the problem that the traditional minimum entropy estimation method does not consider the influence of environmental factors and other multi-source data on the quality of physical entropy source random number. A multi-source information fusion physical entropy source security evaluation method is provided, comprising the following steps: obtaining physical entropy source data and processing to obtain a multi-source fusion data set; performing feature extraction on the multi-source fusion data set through a variational autoencoder network, and outputting a latent representation as the data after feature extraction; inputting the latent representation divided into a test set into a 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 security of the physical entropy source based on the minimum entropy evaluation result. The present application can improve the low accuracy of existing physical entropy source evaluation.

Description

Technical Field

[0001] This invention belongs to the fields of information security and random number generator technology, specifically relating to a method for assessing the security of physical entropy sources through multi-source information fusion. Background Technology

[0002] Random number generators are widely used in spread spectrum communication, scientific computing, numerical analysis, and authentication, especially in cryptographic applications and 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 number sequences with approximate randomness and potential predictability. Unlike pseudo-random number generators, true random number generators generate truly random number sequences based on unpredictable physical random phenomena or processes (i.e., physical entropy sources), thus offering better security. Therefore, highly accurate minimum entropy estimation methods are crucial 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., time series of random numbers) and do not take into account the impact of multi-source data, such as environmental factors (e.g., temperature and voltage fluctuations), on the quality of random numbers. Physical entropy sources (e.g., thermal noise and electronic noise) are very sensitive to changes in temperature, power supply voltage, etc., and these 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 in the prior art, the present invention provides a method for assessing the security of physical entropy sources by fusing multi-source information.

[0005] This invention employs the following technical solution: a method for assessing the security of physical entropy sources through multi-source information fusion, comprising the following steps:

[0006] The process involves acquiring and processing physical entropy source data to obtain a multi-source fusion dataset. This dataset includes the raw data of the external temperature, power supply voltage, and output random number sequence of the physical entropy source. A variational autoencoder network is used to extract features from the multi-source fusion dataset, outputting latent representations as the extracted data. These latent representations are then divided into training, validation, and test sets. An initial hybrid deep learning neural network model is constructed, and the model is trained and validated based on the latent representations divided into training and validation sets. The latent representations divided into test sets are input into the validated hybrid deep learning neural network model to obtain prediction results for the physical entropy sources. Global and local prediction probabilities are calculated based on these prediction results. A minimum entropy evaluation result is obtained based on the global and local prediction probabilities. The safety of the physical entropy source is determined 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 latent representation output by the variational autoencoder network includes the external temperature, power supply voltage, and output random number sequence of the physical entropy source after feature extraction.

[0008] Preferably, in the latent representation output by the variational autoencoder network, every 10 consecutive adjacent numbers are used as a sequence. The first of two adjacent sequences is used as the input sequence of the hybrid deep learning neural network model, and the second sequence is used as the output prediction of the first sequence in the hybrid deep learning neural network model, so as to train and validate the hybrid deep learning neural network model.

[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 CNN layers, where the input sequence is fed into the CNN layer for spatial scale feature extraction; the LSTM network includes LSTM layers, where the output of the LSTM layers corresponds to the hidden state at each time step in the input sequence; the TPA attention mechanism layer includes CNN layers and an evaluation scoring function, used for time series analysis of long sequences; and 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. Then, all hidden states are segmented, and all information is divided into hidden state H and current state h. t ;

[0013] The hidden state information is input to the TPA (Time-Attention Allocation) mechanism layer to extract long-term temporal patterns from the input information. A CNN filter is applied to the row vectors of the hidden state H. A one-dimensional convolutional CNN layer within the TPA mechanism layer performs convolution operations on each row vector, selects the ReLU activation function, and outputs H. C As a feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer serves 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 magnitude of the weights, and the current LSTM output state h is determined. tThe input is repeatedly fed into the Multiply layer through the RepeatVector layer to calculate its relationship with H. C The product of each row of the matrix is ​​input, and the result of the product of each row is summed on the time index axis. After summation, the result is normalized using the Sigmoid activation function. An attention weight 'a' is then calculated for each row, representing H. C The h to be predicted in each row of the matrix t The intensity of the effect; based on each time series, h t The strength of the influence is determined by weighting and summing the weights of each row to obtain the overall time attention v. t , will v t Compared with the current state h t The output h′ of the attention mechanism layer is obtained after merging through the Add layer. t .

[0014] Preferably, 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 among them 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 as follows:

[0016] H min =-log2(max(P) global ,P local ))

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

[0018] Preferably, the formula for determining the safety of a physical entropy source based on the minimum entropy assessment result is as follows:

[0019] H min >0.99×M

[0020] In the formula, M = 1, 2, 3...

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] The physical entropy source security assessment based on multi-source information fusion provided by this invention can improve the low accuracy of existing physical entropy source assessments. If the minimum entropy estimation result meets the specified size, the random numbers generated by the entropy source are considered unpredictable, thereby ensuring information security. The minimum entropy assessment accuracy of this invention is superior to existing statistical and deep learning-based assessment methods. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the minimum entropy estimation process for the physical entropy source of this invention;

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

[0026] Figure 3 This is a diagram of the hybrid deep learning model structure (CNN-TPA-LSTM) of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should fall within the scope of the technical content disclosed in 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] This invention provides an embodiment:

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

[0031] In this embodiment, the steps for processing the physical entropy source data include: normalizing the external temperature and power supply voltage data of the physical entropy source, standardizing the random number sequence in the form of M-bit integers, 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 dataset, using 40% of the entire dataset to build the training set of the model, 10% to evaluate the model but not update the validation set, and 50% as the test set.

[0032] The entire dataset is then fed into a feature extraction variational autoencoder model with a latent dimension of 10 and an intermediate layer dimension of 32. The learner Adam is manually set to a learning rate of 0.001, and the loss function is a linear combination of cross-entropy loss and KL divergence. The model is then trained using an encoder (LSTM network) and a decoder (LSTM network), and gradients are calculated to update the model until the loss no longer decreases, at which point the latent 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 cross-entropy loss and KL divergence. The latent representation output by the variational autoencoder network includes the external temperature, supply voltage, and output random number sequence of the physical entropy source after feature extraction.

[0034] Hybrid deep learning neural network models include CNN networks, LSTM networks, TPA attention mechanism layers, and fully connected networks;

[0035] The CNN network includes CNN layers, where the input sequence is fed into the CNN layer for spatial scale feature extraction; the LSTM network includes LSTM layers, where the output of the LSTM layers corresponds to the hidden state at each time step in the input sequence; the TPA attention mechanism layer includes CNN layers and an evaluation scoring function, used for time series analysis of long sequences; and 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] In the latent representation of the variational autoencoder network output, every 10 consecutive adjacent numbers are grouped into a sequence. The first of two adjacent sequences is used as the input sequence of the hybrid deep learning neural network model, and the second sequence is used as the output prediction (label) of the first sequence in the hybrid deep learning neural network model. Next, the sequence is shifted one position to the right 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, which are used to train and validate the hybrid deep learning neural network model.

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

[0039] In the CNN layer of the CNN network, local features are further extracted from the latent representation input. The number of filters is k=256, the kernel size is 9, and Rel is used as the activation function.

[0040] In the LSTM layer, the time series fed by the 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 hidden states are segmented, and all information is divided into hidden states H = (h1, h2, ..., 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 from the input information. The model's prediction performance is improved by applying CNN filters to the row vectors of H. A one-dimensional convolutional CNN layer within the TPA temporal attention mechanism layer performs convolution operations on each row vector. This one-dimensional convolutional CNN layer has 256 one-dimensional convolutional kernels, each with a length of 10, selects the ReLU activation function, and outputs H. C As a feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer serves as a scoring function to determine the weights of each temporal pattern in the output of the one-dimensional convolutional CNN layer. Attention is calculated based on the weights, and the current LSTM output state h is determined. t The input is repeatedly fed into the Multiply layer through the RepeatVector layer to calculate its relationship with H. C The product of each row of the matrix is ​​input, and the result of the product of each row is summed on the time index axis. After summation, the result is normalized using the Sigmoid activation function. An attention weight 'a' is then calculated for each row, representing H. C The h to be predicted in each row of the matrix t The intensity of the effect. Based on each time series, h... t The strength of the influence is determined by weighting and summing the weights of each row to obtain the overall time attention v. t , will v t Compared with the current state h t The output h′ of the attention mechanism layer is obtained after merging through the Add layer. t .

[0042] In the fully connected layers of a fully connected network, the Softmax activation function is used to map the linearly transformed output vector to a set of probability distributions. Assume the linear output of the output layer is z = (z0, z1, ... zz). s ), where s is the size of the sample space, equal to the output size of the fully connected layer. The Softmax activation function is defined as... We obtain, where y i This represents the probability corresponding to the i-th output category, such that the sum of all output probabilities equals 1. In the formula, i = 1, 2, 3, ...

[0043] Furthermore, a manually set optimizer, Adam, with a learning rate of 0.0005, is introduced to minimize the objective function during training. During training and validation, a batch of 256 samples is input into the model, and training is performed using a cross-entropy loss function. The maximum number of training epochs is set to 1000, and an early stopping mechanism is implemented: training stops when the loss no longer decreases after 7 consecutive epochs. The validation set loss and corresponding training weights are recorded during training, and the training weights with the minimum validation error are saved to evaluate the final training weights for the corresponding test set.

[0044] After completing model training, input the test dataset and then follow the same method used in 90B for minimum entropy estimation to calculate the global and local predictability probabilities.

[0045] Given a discrete, independent random variable X, its possible values ​​are set as A = {x1, x2, ..., xn}. k The 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 quantity that can take the values ​​of each element in set A, with each value having a corresponding probability. For any x i ∈A, P r (X=x i ) = p i , i = 1, ..., k, where p i Corresponding to y above i The definition of minimum entropy is:

[0046] Deep learning models predict the next sample in a sequence based on specific statistical properties of previous samples and provide evaluation results based on the predictions. The trained model is tested on an untrained dataset to estimate the global and local probabilities of successful predictions at a 99% confidence level, and the minimum entropy is calculated from the predictions.

[0047] (1) Global prediction probability: Assume N is the sample size, N T Let be the number of correct predictions, and r be the number of consecutive correct predictions. Prediction accuracy P. r The formula is given as follows:

[0048] At a 99% confidence level, the global prediction probability P global The calculation is as follows:

[0049]

[0050] (2) Local prediction probability: The local prediction probability is calculated iteratively 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 as follows:

[0054] H min =-log2(max(P) global ,P local ))

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

[0056] The formula for determining the safety of a physical entropy source based on the minimum entropy assessment results is as follows:

[0057] H min >0.99×M

[0058] In the formula, M = 1, 2, 3...

[0059] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assessing the security of physical entropy sources through multi-source information fusion, characterized in that, Includes the following steps: The physical entropy source data is acquired and processed to obtain a multi-source fusion dataset; the physical entropy source data includes the original data of the external temperature, power supply voltage and output random number sequence of the physical entropy source. The multi-source fusion dataset is subjected to feature extraction using a variational autoencoder network, which outputs a latent representation as the data after feature extraction. The latent representation is then divided into a training set, a validation set, and a test set. 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 latent representation output by the variational autoencoder network includes the external temperature, power supply voltage, and output random number sequence of the physical entropy source after feature extraction. An initial hybrid deep learning neural network model was built, and the model was trained and validated based on the latent representations divided into training and validation sets. The hybrid deep learning neural network model includes a CNN network, an LSTM network, a TPA attention mechanism layer, and a fully connected network. The CNN network consists of CNN layers, where the input sequence is fed into the CNN layer for spatial scale feature extraction. The LSTM network consists of LSTM layers, whose outputs correspond to the hidden states at each time step in the input sequence. The TPA attention mechanism layer includes CNN layers and an evaluation scoring function for long-sequence time series analysis. The fully connected network obtains the corresponding output probabilities through the Softmax activation function. The CNN network... The output of the layer 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. In the latent representation output by the variational autoencoder network, every 10 consecutive adjacent numbers are grouped into a sequence. The first of two adjacent sequences is used as the input sequence of the hybrid deep learning neural network model, and the second sequence is used as the output prediction of the first sequence in the hybrid deep learning neural network model. This process is used to train and validate the hybrid deep learning neural network model. The latent representations divided into a test set are then input into the validated hybrid deep learning neural network model to obtain the prediction results of the physical entropy source. Calculate the global prediction probability and the local prediction probability based on the prediction results; The minimum entropy evaluation result is obtained based on the global prediction probability and the local prediction probability; The safety of a physical entropy source is determined based on the minimum entropy assessment results.

2. The method for assessing the security of physical entropy sources through multi-source information fusion according to claim 1, characterized in that: In the LSTM layer, the time series fed by the CNN network is received, and the output is the hidden state corresponding to each time step in the input sequence. Then, all hidden states are segmented, and all information is divided into hidden states. and current state ; Hidden state information is input to the TPA (Time Attention) mechanism layer to extract long-term temporal patterns from the input information and process the hidden state. The row vectors are filtered using a CNN filter. A one-dimensional convolutional CNN layer within the TPA (Time Attention Allocation) mechanism layer performs convolution operations on each row vector, selecting the ReLU activation function and outputting the result. As a feature representation matrix, a fully connected layer in the TPA temporal attention mechanism layer serves 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 magnitude of the weights, and the current LSTM output state is determined. The input is repeatedly passed to the Multiply layer via the RepeatVector layer for calculating its sum. The product of each row of the matrix is ​​input, and the result of the product of each row is summed on the time index axis. After summation, the result is normalized using the Sigmoid activation function, and an attention weight corresponding to each row is calculated. Its representative The matrix to be predicted in each row The intensity of the effect; based on each time series The strength of the influence is determined by weighting and summing the weights of each row to obtain the overall time attention. ,Will Compared with the current state The output of the attention mechanism layer is obtained after merging through the Add layer. .

3. The method for assessing the security of physical entropy sources through multi-source information fusion according to claim 1, characterized in that: In the fully connected layer of a fully connected network, using The activation function maps the linearly transformed output vector to a set of probability distributions and selects the one with the highest probability as the output.

4. The method for assessing the security of physical entropy sources through multi-source information fusion according to claim 1, characterized in that: The minimum entropy evaluation result obtained based on the global prediction probability and the local prediction probability is as follows: In the formula, The result is the minimum entropy evaluation result. It is the maximum of the global prediction probability and the local prediction probability. For global prediction probability, This represents the local prediction probability.

5. The method for assessing the security of physical entropy sources through multi-source information fusion according to claim 4, characterized in that: The formula for determining the safety of a physical entropy source based on the minimum entropy assessment results is as follows: In the formula, .

Images