Cryptographic chip side channel analysis method based on multi-dimensional fusion deep learning

Abstract

The application discloses a password chip side channel analysis method based on multi-dimensional fusion deep learning, which comprises the following steps: collecting side channel leakage curves, recording corresponding plaintext, ciphertext or operation state information, and forming a side channel data set; dividing the side channel data set into a training set and a test set; constructing a multi-dimensional fusion convolution residual dendritic network; training the multi-dimensional fusion convolution residual dendritic network; inputting the side channel leakage curves in the test set into the trained multi-dimensional fusion convolution residual dendritic network, obtaining the probability distribution corresponding to each key candidate, and cumulatively sorting the probability results of the multiple side channel leakage curves to recover the key of the target cryptographic device. The application extracts leakage signal features from different sampling scales through multi-scale convolution branches and multi-dimensional fusion mechanisms, can simultaneously utilize local leakage changes and cross-sampling point combination information, reduces feature omission caused by a single convolution scale, and improves the side channel leakage signal feature extraction capability.

Description

Technical Field

[0001] This invention relates to the field of transmission, and in particular to a method for side-channel analysis of cryptographic chips based on multi-dimensional fusion deep learning. Background Technology

[0002] With the widespread deployment of IoT terminals, embedded cryptographic modules, smart edge devices, and various security chips, cryptographic technology has become a crucial infrastructure for ensuring data security, identity authentication, and communication security. While cryptographic algorithms possess high mathematical security, cryptographic devices inevitably experience physical information leaks during actual operation, such as power consumption, electromagnetic radiation, execution time, and cache access behavior, thus creating side-channel information. Analysts can collect and analyze this side-channel leakage information to infer intermediate variables within the cryptographic algorithm and ultimately recover the key. Therefore, the physical implementation security of cryptographic devices has become an important research direction in the field of cryptographic engineering.

[0003] Side-channel analysis (SSA) is a technique that utilizes physical leaks generated during the operation of cryptographic devices for analysis. Depending on the analysis method, traditional SSA can be categorized into simple power analysis, differential power analysis, correlated power analysis, and template analysis. These methods typically rely on the analyst's prior knowledge of the leakage model and the characteristics of the target device, requiring the construction of specific statistical models or manual feature extraction. The analysis effectiveness is easily affected by noisy environments, device differences, and the accuracy of the leakage model, and often suffers from insufficient generalization ability in complex scenarios.

[0004] With the development of artificial intelligence technology, deep learning side-channel analysis has gradually become an important direction in side-channel analysis research. Deep learning methods can automatically learn feature representations directly from the original side-channel leakage curves, without relying on manually designed features or precise leakage models, thus effectively improving the analysis model's adaptability to noise, clock jitter, and device differences. Currently, deep learning models such as Convolutional Neural Networks (CNN), Residual Networks (ResNet), Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), and Multilayer Perceptrons (MLP) have been widely applied to side-channel analysis tasks of cryptographic algorithms such as AES, achieving good key recovery results.

[0005] However, existing deep learning side-channel analysis methods still have certain limitations. First, many models primarily focus on the final classification accuracy or key recovery success rate, while giving less consideration to the impact of model training efficiency and modeling costs on practical analysis scenarios. When the training sample size is large or the model structure is deep, it often requires long training times and high computational resource overhead. Second, side-channel leakage curves typically contain local transient leakage features, interval correlation features, and cross-sampling point combination features simultaneously. Single-scale convolutional structures are insufficient to fully extract effective leakage information at different time scales, easily leading to the neglect of some key features. Third, as network depth increases, models are prone to problems such as gradient vanishing, decreased training convergence speed, and overfitting, thus affecting model stability and analysis performance. Furthermore, existing models have limited ability to mine the correlations between features at different levels, making it difficult to fully utilize the complementary information between multi-scale features, thus limiting feature representation capabilities.

[0006] Therefore, there is an urgent need for a cryptographic chip side-channel analysis method that can simultaneously utilize leakage features at different time scales, enhance feature association expression capabilities, improve model training efficiency, and maintain high key recovery accuracy, in order to address the shortcomings of existing deep learning side-channel analysis models in terms of feature extraction capabilities, training efficiency, and analysis performance. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a simple and highly accurate method for analyzing the side channel of cryptographic chips based on multidimensional fusion deep learning.

[0008] The technical solution of this invention to solve the above-mentioned technical problems is: a cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, comprising the following steps:

[0009] Step 1: Collect the side channel leakage curve generated when the target cryptographic device executes the cryptographic algorithm, and record the plaintext, ciphertext or operation state information corresponding to the side channel leakage curve to form a side channel dataset;

[0010] Step 2: Preprocess the side-channel dataset and then divide it into training and test sets;

[0011] Step 3: Construct a multidimensional fused convolutional residual dendritic network;

[0012] Step 4: Train a multidimensional fusion convolutional residual dendritic network using the side-channel leakage curves in the training set, so that the multidimensional fusion convolutional residual dendritic network learns the mapping relationship between the side-channel leakage curves and key-related intermediate values.

[0013] Step 5: Input the side-channel leakage curves from the test set into the trained multidimensional fusion convolutional residual dendritic network to obtain the probability distribution corresponding to each key candidate, and accumulate and sort the probability results of multiple side-channel leakage curves to recover the key of the target cryptographic device or output the key candidate sorting result.

[0014] In the above-mentioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, in step 1, the side-channel leakage curve is obtained by a power consumption acquisition device, an electromagnetic probe, or a timing measurement device; the side-channel leakage curve includes at least one of power consumption signal, electromagnetic radiation signal, and clock jitter signal.

[0015] The above-mentioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning includes, in step 2, preprocessing including curve filtering, time window pruning, and standardization.

[0016] The curve screening process involves removing side-channel leakage curves that are abnormally collected or significantly distorted.

[0017] The time window trimming process is as follows: First, the leakage curves of each side channel are synchronized in time. Then, the signal-to-noise ratio (SNR) value of each sampling point of the side channel leakage curve is calculated by the SNR analysis method. The sampling area with an SNR higher than the preset threshold is selected as the point of interest area and the corresponding sampling window is truncated.

[0018] The standardization process involves normalizing the side-channel leakage curve using the mean-variance standardization method.

[0019] The above-mentioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, in step 2, after preprocessing, labels are constructed based on key-related intermediate values ​​during the cryptographic algorithm's operation. These labels are then used to realize a supervised learning mapping relationship between side-channel leakage information and key information. The specific process is as follows:

[0020] Let the first The side channel leakage curve corresponds to the 0th byte of plaintext. The candidate key bytes are Key-related intermediate values ​​are represented as follows:

[0021] ;

[0022] in, For the first The side-channel leakage curve is in the candidate key bytes. Intermediate values ​​related to the key at that time; Indicates AES S-box transformation; This represents the XOR operation;

[0023] A 256-class label system is constructed using an identity leakage model; therefore, the labels are represented as follows:

[0024] ;

[0025] in, For the first The side-channel leakage curve is in the candidate key bytes. The label at the time.

[0026] The above-mentioned cryptographic chip side-channel analysis method based on multidimensional fusion deep learning, in step 3, the multidimensional fusion convolutional residual dendritic network includes the original signal input layer, multi-scale convolutional branches, early fusion layer, residual dendritic network module, late fusion layer and classification output layer;

[0027] The multi-scale convolutional branches include small-scale convolutional branches for extracting local transient leakage features and large-scale convolutional branches for extracting combined leakage features across time windows. The outputs of each convolutional branch are combined into a unified leakage feature representation through an early fusion layer.

[0028] In the early fusion layer, let the feature tensors obtained from the small-scale convolution branches be respectively The feature tensor obtained by large-scale convolution branches Then, additive fusion is represented as:

[0029] ;

[0030] in, This represents the leakage feature representation obtained by additive fusion;

[0031] Subtraction fusion is represented as:

[0032] ;

[0033] in, This represents the leakage feature obtained by subtraction fusion;

[0034] Multiplication fusion is represented as:

[0035] ;

[0036] in, This represents the dot product operation; This represents the leakage feature representation obtained by multiplication fusion;

[0037] Average fusion is expressed as:

[0038] ;

[0039] in, This represents the leakage feature obtained by average fusion;

[0040] Choose one of the four fusion methods to use.

[0041] In the above-mentioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, step 3 includes a residual dendritic network module comprising a first residual dendritic network, a second residual dendritic network, a third residual dendritic network, and a fourth residual dendritic network.

[0042] The computation process in the first residual dendritic network is represented as follows:

[0043] ;

[0044] in, Represents the Hadamard product; This represents the output of the first residual dendritic network, and As input to the second residual dendritic network; This represents the weight matrix of the output of the first residual dendritic network; This represents the original input of the residual dendritic network module, which is the leakage feature representation of the output of the early fusion layer;

[0045] The computation process in the second residual dendritic network is represented as follows:

[0046] ;

[0047] in, This represents the output of the second residual dendritic network, and As input to the third and fourth residual dendritic networks; The weight matrix represents the output of the second residual dendritic network;

[0048] The computation process in the third residual dendritic network is represented as follows:

[0049] ;

[0050] in, This represents the output of the third residual dendritic network; The weight matrix represents the output of the third residual dendritic network;

[0051] The computation process in the fourth residual dendritic network is represented as follows:

[0052] ;

[0053] in, This represents the output of the fourth residual dendritic network; This represents the weight matrix of the output of the fourth residual dendritic network.

[0054] In the aforementioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, step 3 involves the later fusion layer fusing the outputs of the third and fourth residual dendritic networks using an additive fusion method. The later fusion layer outputs a feature vector. , :Will Rewritten in vector form, we get , This represents the feature vector. The first in Dimensional features;

[0055] The output probability of the classification output layer is represented as follows:

[0056] ;

[0057] in, Indicates the intermediate value category number related to the key. This indicates that the classification output layer corresponds to the first... Output values ​​for each category; This indicates that the classification output layer corresponds to the first... Output values ​​for each category; Indicates category index, Indicates the first The predicted probability that the side channel leakage curve belongs to class c; Indicates the total number of categories. and equal.

[0058] In the aforementioned method for side-channel analysis of cryptographic chips based on multidimensional fusion deep learning, in step 4, during the training process, the multidimensional fusion convolutional residual dendritic network trains the first... The true label of the side channel leakage curve is The number of training samples is Then the cross-entropy loss function during the training process is expressed as:

[0059] ;

[0060] in, Represents cross-entropy loss; Indicates the first The side channel leakage curve belongs to The predicted probability.

[0061] In the above-mentioned cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning, step 5 involves inputting the side-channel leakage curves from the test set into the trained multi-dimensional fusion convolutional residual dendritic network to obtain the probability distribution of each side-channel leakage curve under each key-related intermediate value category.

[0062] For candidate key bytes The intermediate value tag is calculated based on the plaintext and candidate key, and the logarithmic probability of multiple side-channel leakage curves is accumulated; candidate key bytes The cumulative log probability is expressed as:

[0063] ;

[0064] in, Represents candidate key bytes The cumulative log probability; Indicates the first Side-channel leakage curve in candidate key bytes The corresponding intermediate value label;

[0065] After calculating the cumulative logarithmic probability of all candidate keys, they are sorted from highest to lowest value. The candidate key ranked first is used as the recovery result or as a candidate for subsequent multi-byte key recovery.

[0066] The beneficial effects of this invention are as follows:

[0067] 1. This invention extracts leakage signal features from different sampling scales through multi-scale convolutional branching and multi-dimensional fusion mechanisms. It can simultaneously utilize local leakage changes and cross-sampling point combination information, reduce feature omissions caused by a single convolutional scale, and improve the side-channel leakage signal feature extraction capability.

[0068] 2. This invention introduces a multidimensional fusion convolutional residual dendritic network, which combines the residual concept with dendritic nonlinear expression, helping to alleviate the degradation and overfitting problems caused by network deepening. Experimental results show that the multidimensional fusion convolutional residual dendritic network proposed in this invention has higher modeling efficiency and faster training convergence speed than existing deep learning side-channel analysis models, improving the training stability and fitting speed of deep learning side-channel analysis models.

[0069] 3. This invention utilizes a multi-dimensional fusion mechanism that combines early and late fusion, and selects a fusion combination suitable for expressing side-channel leakage features from various fusion methods. Experimental results show that this invention can achieve higher analysis accuracy and better key recovery performance on typical AES side-channel analysis datasets.

[0070] 4. This invention modularizes data acquisition, preprocessing, model training, analysis and inference, and key evaluation, and can output evaluation results such as key candidate ranking, guessing entropy, and analysis success rate. It is suitable for side-channel security detection of cryptographic chips, embedded terminals, and IoT devices. Attached Figure Description

[0071] Figure 1 This is the overall flowchart of the present invention.

[0072] Figure 2 This is the original power consumption waveform diagram of the present invention.

[0073] Figure 3 This is a schematic diagram of the signal-to-noise ratio analysis of the present invention.

[0074] Figure 4 This is a comparison chart of the model training accuracy of this invention and other methods.

[0075] Figure 5 This is a comparison chart of the model training loss values ​​of this invention and other methods. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] like Figure 1 As shown, the cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning includes the following steps:

[0078] Step 1: Collect the side-channel leakage curve generated when the target cryptographic device executes the cryptographic algorithm, and record the plaintext, ciphertext or operation state information corresponding to the side-channel leakage curve to form a side-channel dataset.

[0079] During the execution of cryptographic operations such as AES-128, the target cryptographic device generates power consumption, electromagnetic radiation, or time leakage information. The side channel acquisition platform synchronously records the side channel leakage curve and the corresponding plaintext, ciphertext, or operation status information. The side channel leakage curve includes at least one of the following signals: power consumption signal, electromagnetic radiation signal, and clock jitter signal.

[0080] Step 2: Preprocess the side-channel dataset and then divide it into training and test sets.

[0081] Preprocessing includes curve filtering, time window trimming, and standardization.

[0082] The curve screening process involves removing side-channel leakage curves that are abnormally collected or significantly distorted.

[0083] The time window trimming process is as follows: First, the leakage curves of each side channel are synchronized in time. Then, the signal-to-noise ratio (SNR) value of each sampling point of the side channel leakage curve is calculated by the SNR analysis method. The sampling area with an SNR higher than the preset threshold is selected as the point of interest area and the corresponding sampling window is truncated.

[0084] The standardization process involves normalizing the side-channel leakage curve using the mean-variance standardization method.

[0085] Taking the AES-128 power consumption curve acquired by the CW308T_STM32F3 target board as an example, the leakage intensity distribution curve is obtained by calculating the signal-to-noise ratio (SNR) values ​​at different sampling points, and the target leakage interval is determined based on the SNR peak position. When the SNR of multiple consecutive sampling points exceeds a preset threshold, the corresponding sampling area is taken as the point of interest region, thereby reducing the impact of irrelevant sampling points on the training process. In this invention, the curve interval of the model to be input is determined based on the SNR peak, a manual window, or a template window to reduce the interference of irrelevant sampling points on model training.

[0086] For the AES-128 software implementation, the output of the first round S-box is selected as the analysis point, and the output of the first round S-box of the AES-128 algorithm is used as the intermediate value related to the key. A supervised learning mapping relationship between side-channel leakage information and key information is achieved by constructing tags. The specific process is as follows:

[0087] Let the first The side channel leakage curve corresponds to the 0th byte of plaintext. The candidate key bytes are Key-related intermediate values ​​are represented as follows:

[0088] ;

[0089] in, For the first The side-channel leakage curve is in the candidate key bytes. Intermediate values ​​related to the key at that time; Indicates AES S-box transformation; Represents the XOR operation;

[0090] A 256-class label system is constructed using an identity leakage model; therefore, the labels are represented as follows:

[0091] ;

[0092] in, For the first The side-channel leakage curve is in the candidate key bytes. The label at the time.

[0093] Compared to Hamming weight or Hamming distance models, the identity leakage model preserves the complete category information of the S-box output, avoiding information loss caused by compressing 256 intermediate values ​​into fewer categories.

[0094] Step 3: Construct a multidimensional fused convolutional residual dendritic network.

[0095] The multidimensional fusion convolutional residual dendritic network includes the original signal input layer, multi-scale convolutional branches, early fusion layer, residual dendritic network module, late fusion layer and classification output layer;

[0096] The multi-scale convolutional branches include at least one small-scale convolutional branch for extracting local transient leakage features and at least one large-scale convolutional branch for extracting combined leakage features across time windows. The outputs of each convolutional branch are combined into a unified leakage feature representation through an early fusion layer.

[0097] In the early fusion layer, let the feature tensors obtained from the small-scale convolution branches be respectively The feature tensor obtained by large-scale convolution branches Then, additive fusion is represented as:

[0098] ;

[0099] in, This represents the leakage feature representation obtained by additive fusion;

[0100] Subtraction fusion is represented as:

[0101] ;

[0102] in, This represents the leakage feature obtained by subtraction fusion;

[0103] Multiplication fusion is represented as:

[0104] ;

[0105] in, This represents the dot product operation; This represents the leakage feature representation obtained by multiplication fusion;

[0106] Average fusion is expressed as:

[0107] ;

[0108] in, This represents the leakage feature obtained by average fusion;

[0109] Choose one of the four fusion methods to use.

[0110] This invention employs a multi-dimensional fusion mechanism that combines early fusion and late fusion. Early fusion is used to fuse features output from convolutional branches at different scales, while late fusion is used to further fuse high-level features output from residual dendritic network modules.

[0111] This invention introduces a residual dendritic network module to improve the nonlinear expressive power and training stability of the model. The residual dendritic network module includes various residual block forms, including ResDD-X (Residual Dendritic Network - Original Input Structure), ResDD-WX (Residual Dendritic Network - Weight and Original Input Fusion Structure), ResDD-A (Residual Dendritic Network - Previous Layer Feature Structure Structure), and ResDD-WA (Residual Dendritic Network - Weight and Previous Layer Feature Fusion Structure Structure). Different residual block forms combine feature mapping relationships in different ways to enhance the nonlinear expressive power of the network. Taking the ResDD-WX (Residual Dendritic Network - Weight and Original Input Fusion Structure) used in this invention as an example, the residual dendritic network module includes a first residual dendritic network, a second residual dendritic network, a third residual dendritic network, and a fourth residual dendritic network.

[0112] The computation process in the first residual dendritic network is represented as follows:

[0113] ;

[0114] in, Represents the Hadamard product; This represents the output of the first residual dendritic network, and As input to the second residual dendritic network; This represents the weight matrix of the output of the first residual dendritic network; This represents the original input of the residual dendritic network module, which is the leakage feature representation of the output of the early fusion layer;

[0115] The computation process in the second residual dendritic network is represented as follows:

[0116] ;

[0117] in, This represents the output of the second residual dendritic network, and As input to the third and fourth residual dendritic networks; The weight matrix represents the output of the second residual dendritic network;

[0118] The computation process in the third residual dendritic network is represented as follows:

[0119] ;

[0120] in, This represents the output of the third residual dendritic network; The weight matrix represents the output of the third residual dendritic network;

[0121] The computation process in the fourth residual dendritic network is represented as follows:

[0122] ;

[0123] in, This represents the output of the fourth residual dendritic network; This represents the weight matrix of the output of the fourth residual dendritic network.

[0124] By introducing a residual dendritic network module, the model can express the combination relationships between different leaked features in nonlinear expansion and alleviate the degradation problem when the network deepens. The Hadamard product is used to establish cross-combination relationships between different features, enabling the model to generate high-order nonlinear feature terms; residual connections are used to preserve original feature information and achieve direct gradient propagation, thereby alleviating the gradient vanishing and network degradation problems during deep network training. Through the combination of dendritic nonlinear mapping and residual connections, the model can simultaneously achieve strong feature representation ability and training stability.

[0125] The late-stage fusion layer uses an additive fusion method to fuse the outputs of the third and fourth residual dendritic networks. The late-stage fusion layer outputs a feature vector. , :Will Rewritten in vector form, we get , This represents the feature vector. The first in Dimensional features;

[0126] The output probability of the classification output layer is represented as follows:

[0127] ;

[0128] in, Indicates the intermediate value category number related to the key. This indicates that the classification output layer corresponds to the first... Output values ​​for each category; This indicates that the classification output layer corresponds to the first... Output values ​​for each category; Indicates category index, Indicates the first The predicted probability that the side channel leakage curve belongs to class c; Indicates the total number of categories. and equal.

[0129] Step 4: Train a multidimensional fused convolutional residual dendritic network using the side-channel leakage curves in the training set, so that the multidimensional fused convolutional residual dendritic network learns the mapping relationship between the side-channel leakage curves and key-related intermediate values.

[0130] During the model training phase, the Adam optimizer is used for parameter updates, the Softmax function is used for the classification output layer, and the cross-entropy loss function is used for the loss function.

[0131] During training, the multidimensional fused convolutional residual dendritic network is used for the first... The true label of the side channel leakage curve is The number of training samples is Then the cross-entropy loss function during the training process is expressed as:

[0132] ;

[0133] in, Represents cross-entropy loss; Indicates the first The side channel leakage curve belongs to The predicted probability.

[0134] One example of the hyperparameters used in this invention is as follows: learning rate of 0.0001, batch size of 128, and number of training rounds of 400.

[0135] During training, the accuracy of the training set, the accuracy of the test set, and the loss value can be monitored. The model parameters are saved when the validation metrics converge or the preset number of rounds is reached.

[0136] Step 5: Input the side-channel leakage curves from the test set into the trained multidimensional fusion convolutional residual dendritic network to obtain the probability distribution corresponding to each key candidate, and accumulate and sort the probability results of multiple side-channel leakage curves to recover the key of the target cryptographic device or output the key candidate sorting result.

[0137] The side-channel leakage curves in the test set are input into the trained multidimensional fusion convolutional residual dendritic network to obtain the probability distribution of each side-channel leakage curve under each key-related intermediate value category.

[0138] For candidate key bytes The intermediate value tag is calculated based on the plaintext and candidate key, and the logarithmic probability of multiple side-channel leakage curves is accumulated; candidate key bytes The cumulative log probability is expressed as:

[0139] ;

[0140] in, Represents candidate key bytes The cumulative log probability; Indicates the first Side-channel leakage curve in candidate key bytes The corresponding intermediate value label;

[0141] After calculating the cumulative logarithmic probability of all candidate keys, they are sorted from highest to lowest value. The candidate key ranked first is used as the recovery result or as a candidate for subsequent multi-byte key recovery.

[0142] In the performance evaluation phase, the experiment used PGE (Probabilistic Guessing Entropy) and single-trajectory analysis accuracy to evaluate the model's performance. Let the true key be at the... The ranking of the candidates after the second analysis is as follows: Then we can guess that entropy can be expressed as the formula:

[0143] ;

[0144] in, Indicates the first The entropy guessed in the sub-analysis; For mathematical expectation;

[0145] when When the accuracy gradually approaches 0 or the real key consistently ranks first, it indicates that the model can recover the target key with fewer side-channel leakage curves. Single-trajectory analysis accuracy is used to evaluate the model's ability to correctly classify intermediate values ​​related to the real key when given a single side-channel leakage curve input.

[0146] Figure 2 The original power consumption waveform is shown below. Figure 3 This is a schematic diagram of signal-to-noise ratio analysis. Figure 2 and Figure 3 Used to locate critical leakage areas related to sensitive operations. Figure 2 The image shows the original power consumption waveform, with the horizontal axis representing the sampling point number and the vertical axis representing the signal amplitude. Figure 3 The corresponding signal-to-noise ratio curve is shown, with the horizontal axis representing the sampling point number and the vertical axis representing the signal-to-noise ratio value. Figure 2 and Figure 3 The red dashed line in the diagram indicates the range of signal-to-noise ratio peaks. This range corresponds to the most sensitive operation moment in the encryption process and contains the most key-related information. Subsequent feature extraction and model training will be mainly based on data within this range.

[0147] To verify the effectiveness of the proposed multidimensional fusion convolutional residual dendritic network in the side-channel analysis task of cryptographic chips, the method of this invention was compared with different deep learning side-channel analysis methods under the same experimental conditions. Specifically, Convolutional Neural Network (CNN) was used as comparison method 2, Multilayer Perceptron (MLP) as comparison method 3, and Dendritic Network (DD) as comparison method 1. The method of this invention is a multidimensional fusion convolutional residual dendritic network.

[0148] Figure 4 and Figure 5 This demonstrates the changes in accuracy and loss during the training process using different methods. Figure 4 Accuracy variation curves for different training methods. Figure 5 Training loss variation curves for different methods. From Figure 4 It can be seen that the accuracy of each method model gradually improves with the increase of training rounds, among which the method of this invention can achieve a higher accuracy in fewer training rounds; Figure 5 It can be seen that the loss value decreases rapidly and remains at a low level during the training process using the method of the present invention.

[0149] Experimental results show that this invention improves the model's ability to express side-channel leakage features through multi-scale convolutional feature extraction, multi-dimensional fusion mechanism, and residual dendritic network module, enabling the model to learn the mapping relationship between the side-channel leakage curve and key-related intermediate values ​​more quickly, thereby improving the model training efficiency.

Claims

1. A method for side-channel analysis of cryptographic chips based on multi-dimensional fusion deep learning, characterized in that, Includes the following steps: Step 1: Collect the side channel leakage curve generated when the target cryptographic device executes the cryptographic algorithm, and record the plaintext, ciphertext or operation state information corresponding to the side channel leakage curve to form a side channel dataset; Step 2: Preprocess the side-channel dataset and then divide it into training and test sets; Step 3: Construct a multidimensional fused convolutional residual dendritic network; Step 4: Train a multidimensional fusion convolutional residual dendritic network using the side-channel leakage curves in the training set, so that the multidimensional fusion convolutional residual dendritic network learns the mapping relationship between the side-channel leakage curves and key-related intermediate values. Step 5: Input the side-channel leakage curves from the test set into the trained multidimensional fusion convolutional residual dendritic network to obtain the probability distribution corresponding to each key candidate, and accumulate and sort the probability results of multiple side-channel leakage curves to recover the key of the target cryptographic device or output the key candidate sorting result.

2. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 1, characterized in that, In step 1, the side channel leakage curve is obtained by a power consumption acquisition device, an electromagnetic probe, or a timing measurement device; the side channel leakage curve includes at least one of the following signals: power consumption signal, electromagnetic radiation signal, and clock jitter signal.

3. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 1, characterized in that, In step 2, the preprocessing includes curve filtering, time window clipping, and standardization. The curve screening process involves removing side-channel leakage curves that are abnormally collected or significantly distorted. The time window trimming process is as follows: First, the leakage curves of each side channel are synchronized in time. Then, the signal-to-noise ratio (SNR) value of each sampling point of the side channel leakage curve is calculated by the SNR analysis method. The sampling area with an SNR higher than the preset threshold is selected as the point of interest area and the corresponding sampling window is truncated. The standardization process involves normalizing the side-channel leakage curve using the mean-variance standardization method.

4. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 1, characterized in that, In step 2, after preprocessing, a label is constructed based on the key-related intermediate values ​​during the operation of the cryptographic algorithm. The label is used to realize the supervised learning mapping relationship between side-channel leakage information and key information. The specific process is as follows: Let the first The side channel leakage curve corresponds to the 0th byte of plaintext. The candidate key bytes are Key-related intermediate values ​​are represented as follows: ； in, For the first The side-channel leakage curve is in the candidate key bytes. Intermediate values ​​related to the key at that time; Indicates AES S-box transformation; Represents the XOR operation; A 256-class label system is constructed using an identity leakage model; therefore, the labels are represented as follows: ； in, For the first The side-channel leakage curve is in the candidate key bytes. The label at the time.

5. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 1, characterized in that, In step 3, the multidimensional fusion convolutional residual dendritic network includes an original signal input layer, a multi-scale convolutional branch, an early fusion layer, a residual dendritic network module, a late fusion layer, and a classification output layer. The multi-scale convolutional branches include small-scale convolutional branches for extracting local transient leakage features and large-scale convolutional branches for extracting combined leakage features across time windows. The outputs of each convolutional branch are combined into a unified leakage feature representation through an early fusion layer. In the early fusion layer, let the feature tensors obtained from the small-scale convolution branches be respectively The feature tensor obtained by large-scale convolution branches Then, additive fusion is represented as: ； in, This represents the leakage feature representation obtained by additive fusion; Subtraction fusion is represented as: ； in, This represents the leakage feature obtained by subtraction fusion; Multiplication fusion is represented as: ； in, This represents the dot product operation; This represents the leakage feature representation obtained by multiplication fusion; Average fusion is expressed as: ； in, This represents the leakage feature obtained by average fusion; Choose one of the four fusion methods to use.

6. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 5, characterized in that, In step 3, the residual dendritic network module includes a first residual dendritic network, a second residual dendritic network, a third residual dendritic network, and a fourth residual dendritic network. The computation process in the first residual dendritic network is represented as follows: ； in, Represents the Hadamard product; This represents the output of the first residual dendritic network, and As input to the second residual dendritic network; This represents the weight matrix of the output of the first residual dendritic network; This represents the original input of the residual dendritic network module, which is the leakage feature representation of the output of the early fusion layer; The computation process in the second residual dendritic network is represented as follows: ； in, This represents the output of the second residual dendritic network, and As input to the third and fourth residual dendritic networks; The weight matrix represents the output of the second residual dendritic network; The computation process in the third residual dendritic network is represented as follows: ； in, This represents the output of the third residual dendritic network; The weight matrix represents the output of the third residual dendritic network; The computation process in the fourth residual dendritic network is represented as follows: ； in, This represents the output of the fourth residual dendritic network; This represents the weight matrix of the output of the fourth residual dendritic network.

7. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 6, characterized in that, In step 3, the late-stage fusion layer uses an additive fusion method to fuse the outputs of the third and fourth residual dendritic networks, and the late-stage fusion layer outputs a feature vector. , :Will Rewritten in vector form, we get , This represents the feature vector. The first in Dimensional features; The output probability of the classification output layer is represented as follows: ； in, Indicates the key-related intermediate value category number, This indicates that the classification output layer corresponds to the first... Output values ​​for each category; This indicates that the classification output layer corresponds to the first... Output values ​​for each category; Indicates category index, Indicates the first The predicted probability that the side channel leakage curve belongs to class c; Indicates the total number of categories. and equal.

8. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 7, characterized in that, In step 4, during the training process, the multidimensional fusion convolutional residual dendritic network is used for the first... The true label of the side channel leakage curve is The number of training samples is Then the cross-entropy loss function during the training process is expressed as: ； in, Represents cross-entropy loss; Indicates the first The side channel leakage curve belongs to The predicted probability.

9. The cryptographic chip side-channel analysis method based on multi-dimensional fusion deep learning according to claim 8, characterized in that, In step 5, the side channel leakage curves in the test set are input into the trained multidimensional fusion convolutional residual dendritic network to obtain the probability distribution of each side channel leakage curve under each key-related intermediate value category. For candidate key bytes The corresponding intermediate value label is calculated based on the plaintext and candidate key, and the logarithmic probability of multiple side channel leakage curves is accumulated. Candidate key bytes The cumulative log probability is expressed as: ； in, Represents candidate key bytes The cumulative log probability; Indicates the first Side-channel leakage curve in candidate key bytes The corresponding intermediate value label; After calculating the cumulative logarithmic probability of all candidate keys, they are sorted from highest to lowest value. The candidate key ranked first is used as the recovery result or as a candidate for subsequent multi-byte key recovery.

Images