A public key cryptography artificial intelligence side channel analysis method and system

Abstract

The application relates to the field of information security cryptography, in particular to a public key cryptography artificial intelligence side channel analysis method and system, wherein the method comprises the following steps: constructing a neural network model for cryptography operation classification for a power consumption channel and an electromagnetic channel, and independently deploying the neural network model to a dual-channel single-chip microcomputer; a host computer respectively pushes cryptography operation waveform data to the dual-channel single-chip microcomputer, executes neural network inference, and outputs inference results containing cryptography operation category labels and corresponding Softmax probability distribution; the host computer takes the Softmax probability distribution as a basic probability assignment, executes double-channel evidence fusion, and obtains a fusion probability distribution corresponding to each cryptography operation waveform data; the target cryptography operation category of each cryptography operation waveform data is judged according to the fusion probability distribution, and the key of a public key cryptography algorithm to be analyzed is recovered. Through the application, the precision and anti-interference ability of key recovery are effectively improved, and low-overhead, high-accuracy portable public key cryptography side channel security analysis can be realized on a resource-limited edge device.

Description

Technical Field

[0001] This invention relates to the field of information security cryptography technology, specifically to a public-key cryptography artificial intelligence side-channel analysis method and system. Background Technology

[0002] Side-channel analysis is a core technical means to evaluate the physical security performance of cryptographic devices. Since the introduction of timing attacks in 1996, public-key cryptography algorithms have become a key target of side-channel analysis due to the differentiated leakage characteristics of modular exponentiation and scalar multiplication operations in terms of power consumption and electromagnetic radiation. Attackers can effectively recover device keys by relying on only a small amount of side-channel waveform data.

[0003] Traditional side-channel analysis methods are mostly based on a single leaked channel, with typical methods including simple power analysis and differential power analysis. However, the leaked information from a single channel suffers from insufficient completeness and is highly susceptible to environmental noise interference, significantly limiting the success rate of such methods in secure cryptographic implementations and low signal-to-noise ratio scenarios. To address this deficiency, the academic community has proposed a multi-channel fusion analysis approach, which effectively improves the effectiveness and robustness of side-channel attacks by simultaneously mining and utilizing leaked information from multiple side channels.

[0004] Currently, deep learning-based multi-channel fusion analysis methods all rely on personal computers or server platforms, resulting in high computational resource requirements and dependence on stable power supplies. This makes them unsuitable for lightweight, scenario-based applications such as on-site safety assessments and portable rapid testing. Deploying this high-precision analysis capability to resource-constrained edge devices (such as microcontrollers) could effectively broaden the application scenarios and scope of side-channel analysis technology. However, edge device deployment still faces several key technical bottlenecks, and no mature and complete solution exists yet. Firstly, the large number of analysis model parameters is mismatched with the limited storage capacity and computing power of microcontrollers. Secondly, the large volume of side-channel waveform data necessitates customized transmission protocols for data interaction. Thirdly, the limited memory resources of microcontrollers prevent the caching and storage of complete waveform data, placing stringent demands on the data processing flow. Fourthly, the multi-channel fusion architecture requires coordinating the model inference and evidence fusion processes of dual-channel microcontrollers, making device coordination and scheduling challenging. Summary of the Invention

[0005] To address these issues, this invention provides a public-key cryptography AI side-channel analysis method and system, aiming to solve the technical problems of existing multi-channel side-channel analysis schemes, such as reliance on high-performance computing devices, poor portability, difficulty in edge deployment, large model parameter volume, high waveform data memory overhead, and unreliable multi-channel data transmission in edge deployment scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] According to a first aspect of the present invention, the present invention provides a public-key cryptography artificial intelligence side-channel analysis method, the method comprising: Neural network models for cryptographic operation classification are constructed for power consumption channels and electromagnetic channels respectively, and neural network compression technology is used to independently deploy each neural network model to the first microcontroller and the second microcontroller. The host computer establishes communication with the first microcontroller and the second microcontroller through two independent serial ports respectively, and uses a packet streaming transmission method with handshake mechanism to push cryptographic operation waveform data of the corresponding channel type to the first microcontroller and the second microcontroller respectively; The first microcontroller and the second microcontroller respectively use the deployed neural network model to perform neural network inference on the cryptographic operation waveform data, output the inference result containing the cryptographic operation category label and the corresponding Softmax probability distribution, and send the inference result back to the host computer; The host computer, based on the Dempster-Shafer evidence theory, uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fusion probability distributions corresponding to each of the cryptographic operation waveform data. The target cryptographic operation category of each cryptographic operation waveform data is determined based on the fusion probability distribution, and the cryptographic operation discrimination result sequence is obtained by splicing them in chronological order. The key of the public key cryptographic algorithm to be analyzed is recovered based on the operation discrimination result sequence.

[0008] Furthermore, the construction of neural network models for cryptographic operation classification, respectively targeting the power consumption channel and the electromagnetic channel, includes: Two independent convolutional neural network models are constructed for the power consumption channel and the electromagnetic channel, respectively. Each convolutional neural network model includes several convolutional layers, batch normalization layers, pooling layers, global average pooling layers, several fully connected layers, and a Softmax output layer. The convolutional neural network model takes the floating-point sampling sequence of a single-segment cryptographic operation waveform as input, automatically extracts side-channel waveform features, completes operation classification, and outputs the cryptographic operation category label and the corresponding Softmax probability distribution of the single-segment cryptographic operation waveform. Each convolutional neural network model is trained independently using dedicated training datasets for the power consumption channel and the electromagnetic channel, respectively, to obtain a power consumption channel cryptographic operation classification model and an electromagnetic channel cryptographic operation classification model.

[0009] Furthermore, the adoption of neural network compression technology, which independently deploys each neural network model to the first microcontroller and the second microcontroller, includes: Knowledge distillation technology is used to lightweight each neural network model to obtain a deployment model adapted for microcontroller operation; An embedded neural network generation tool is used to convert the model files of each deployment model into compilable code, and the compilable code is integrated into the microcontroller projects of the first microcontroller and the second microcontroller, respectively. The first microcontroller deploys the deployment model corresponding to the power consumption channel; the second microcontroller deploys the deployment model corresponding to the electromagnetic channel.

[0010] Furthermore, the packet streaming transmission method with handshake mechanism is implemented using a preset serial communication protocol, including a handshake phase, a data transmission phase, and a summarization phase; During the handshake phase, the host computer sends a handshake frame to the microcontroller, and the microcontroller replies with an acknowledgment frame carrying communication parameters, thus completing the communication link initialization. During the data transmission phase, the host computer first sends a session start frame to carry the total number of cryptographic waveform segments and the algorithm type. Then, for each cryptographic waveform data segment, it sequentially sends a segment start frame, a streaming data frame carrying the cryptographic waveform data, and a segment end frame to the microcontroller. When the microcontroller receives the streaming data frame, it synchronously writes the waveform sampling points contained in the streaming data frame into the neural network input tensor. After receiving the segment end frame, it triggers the neural network inference and sends the inference result back to the host computer. During the aggregation phase, the host computer sends a session end frame to the microcontroller, and the microcontroller resets its own state after receiving the session end frame.

[0011] Furthermore, in the preset serial communication protocol, the frame structure includes a start flag, command word, sequence number, payload length, payload data, and cyclic redundancy check field; The start flag is used to identify the starting position of a single frame of data. The command word is used to distinguish frame types; the frame types include at least one of handshake frames, session start frames, segment start frames, streaming data frames, segment end frames, response frames, and error frames. The sequence number is used to establish the association between request frames and response frames; The payload length is used to identify the number of bytes of payload data; The cyclic redundancy check field is used to identify data transmission errors; The response frame is used to add a status code field at the beginning of the payload data to identify the data interaction status.

[0012] Furthermore, the first microcontroller and the second microcontroller respectively utilize the deployed neural network model to perform neural network inference on the cryptographic waveform data, including: During initialization, the microcontroller pre-allocates activation memory for the neural network inference using a static memory allocation method and obtains the neural network input tensor pointer. The microcontroller firmware runs based on a state machine mechanism. It triggers the transition of working state through command words and responds to valid command instructions. At the same time, it uses DMA circular buffer combined with serial port idle interrupt to receive data. After the interrupt is triggered, it parses the frame structure and completes CRC transmission verification. When the microcontroller receives a streaming data frame, it copies the waveform sampling points within the streaming data frame payload to the corresponding offset positions of the neural network input tensor in situ via memory copying, based on the acquired neural network input tensor pointer, and completes the filling of all waveform sampling points in the single-segment cryptographic operation waveform data frame by frame; after all the single-segment cryptographic operation waveform data has been received and filled, the neural network inference is executed.

[0013] Furthermore, based on the Dempster-Shafer evidence theory, the host computer uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments to perform dual-channel evidence fusion, obtaining the fused probability distributions corresponding to each of the cryptographic operation waveform data, including: Establish an identification framework that includes two categories of cryptographic operations. The two Softmax probability distributions output by the first microcontroller and the second microcontroller are used as the two basic probability values ​​corresponding to the Dempster-Shafer evidence theory. Further, the step of determining the target cryptographic operation category of each of the cryptographic operation waveform data based on the fusion probability distribution, concatenating them in chronological order to obtain a cryptographic operation discrimination result sequence, and recovering the key of the public-key cryptography algorithm to be analyzed based on the operation discrimination result sequence includes: For single-segment cryptographic operation waveform data, the cryptographic operation category with the largest fusion probability distribution is selected as the target cryptographic operation category; Arrange the target cryptographic operation categories corresponding to all the single-segment cryptographic operation waveform data in the order of waveform acquisition time to obtain a cryptographic operation discrimination result sequence; Based on the mapping relationship between the operation type of the public-key cryptography algorithm to be analyzed and the key bits, the sequence of cryptographic operation discrimination results is parsed bit by bit to obtain the key of the public-key cryptography algorithm.

[0014] Furthermore, the public-key cryptography algorithm includes at least one of elliptic curve cryptography, RSA algorithm, and lattice-based key encapsulation mechanism.

[0015] According to a second aspect of the present invention, the present invention provides a public-key cryptography artificial intelligence side-channel analysis system for executing the public-key cryptography artificial intelligence side-channel analysis method according to any one of the first aspects of the present invention, the system comprising a host computer, a first microcontroller, and a second microcontroller; The first microcontroller and the second microcontroller are respectively deployed with neural network models that have been processed by neural network compression technology and perform cryptographic operations to classify power consumption channels and electromagnetic channels; The host computer is used to establish communication with the first microcontroller and the second microcontroller through two independent serial ports respectively, and adopts a packet streaming transmission method with handshake mechanism to push cryptographic operation waveform data of corresponding channel type to the first microcontroller and the second microcontroller respectively. The first microcontroller and the second microcontroller are respectively used to perform neural network inference on the cryptographic operation waveform data using the deployed neural network model, output the inference result containing the cryptographic operation category label and the corresponding Softmax probability distribution, and send the inference result back to the host computer; The host computer is also used to perform dual-channel evidence fusion based on Dempster-Shafer evidence theory, using the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments to obtain the fusion probability distributions corresponding to each of the cryptographic operation waveform data; determine the target cryptographic operation category of each of the cryptographic operation waveform data according to the fusion probability distributions, and concatenate them in chronological order to obtain a cryptographic operation discrimination result sequence; and recover the key of the public-key cryptography algorithm to be analyzed based on the operation discrimination result sequence.

[0016] This invention proposes a public-key cryptography AI side-channel analysis method. It constructs neural network models for cryptographic operation classification, targeting both power consumption and electromagnetic channels respectively. Using neural network compression technology, each neural network model is independently deployed to a first microcontroller and a second microcontroller. A host computer establishes communication with the first and second microcontrollers via two independent serial ports, employing a packet-splitting streaming transmission method with a handshake mechanism to push cryptographic operation waveform data corresponding to the channel type to the first and second microcontrollers respectively. The first and second microcontrollers then utilize their deployed neural network models to perform neural network push operations on the cryptographic operation waveform data. The system outputs an inference result containing cryptographic operation category labels and corresponding Softmax probability distributions, and sends the inference result back to the host computer. Based on Dempster-Shafer evidence theory, the host computer uses the Softmax probability distributions sent back by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fused probability distribution corresponding to each cryptographic operation waveform data. Based on the fused probability distribution, the target cryptographic operation category of each cryptographic operation waveform data is determined, and the cryptographic operation discrimination result sequence is obtained by concatenating the results in chronological order. The key of the public-key cryptographic algorithm to be analyzed is recovered based on the operation discrimination result sequence. This invention effectively improves the accuracy and anti-interference capability of key recovery, enabling low-overhead, high-accuracy portable public-key cryptographic side-channel security analysis on resource-constrained edge devices.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0019] Figure 1 A flowchart illustrating the public-key cryptography AI side-channel analysis method provided in an embodiment of the present invention is shown. Figure 2 A schematic diagram of the structure of the convolutional neural network model provided in an embodiment of the present invention is shown. Figure 3 The diagram shows an electromagnetic side-channel waveform trace generated during the operation of an example of a public-key cryptography algorithm provided in an embodiment of the present invention. Figure 4This diagram illustrates the curves showing the changes in accuracy and loss function during the model distillation training process provided in this embodiment of the invention. Figure 5 This diagram illustrates the frame format and timing of the packet-splitting serial communication protocol provided in an embodiment of the present invention. Figure 6 A schematic diagram illustrating the principle of dual-channel evidence fusion provided in an embodiment of the present invention is shown; Figure 7 The figure shows a comparison curve of the classification confidence of the model provided in the embodiment of the present invention; Figure 8 A schematic diagram of the structure of the public-key cryptography artificial intelligence side-channel analysis system provided in an embodiment of the present invention is shown. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0022] This invention provides a public-key cryptography artificial intelligence side-channel analysis method, such as... Figure 1 As shown, it may include at least the following steps S101~S105: In step S101, neural network models for cryptographic operation classification are constructed for the power consumption channel and the electromagnetic channel, respectively, and neural network compression technology is used to independently deploy each neural network model to the first microcontroller and the second microcontroller.

[0023] First, two independent convolutional neural network models are constructed for the power consumption channel and the electromagnetic channel, respectively. In this embodiment of the invention, one-dimensional convolutional neural network models adapted for cryptographic waveform feature classification are independently constructed for the two types of leakage characteristics: power consumption side channel and electromagnetic side channel. The original training models are simplified and adapted using lightweight neural network compression optimization technology to obtain lightweight deployment models adapted to the computing power, storage, and memory resource constraints of the microcontroller. The independent deployment models corresponding to the two channels are then burned and loaded into two independent microcontroller hardware devices.

[0024] Specifically, the dual-channel convolutional neural network models constructed in the embodiments of the present invention have the same main structure, all consisting of multiple one-dimensional convolutional layers, batch normalization layers, pooling layers, global average pooling layers, multiple fully connected layers, and a Softmax output layer stacked sequentially. The complete network structure is as follows: Figure 2 As shown.

[0025] like Figure 2 As shown, the convolutional neural network model in this embodiment of the invention includes several convolutional layers, batch normalization layers, pooling layers, global average pooling layers, several fully connected layers, and a Softmax output layer. The one-dimensional convolutional layers are used to automatically extract subtle temporal leakage features from the cryptographic operation waveform. The batch normalization layers accelerate model convergence and improve training stability. The pooling layers reduce dimensionality and noise, and compress the size of feature parameters. The global average pooling layer effectively avoids overfitting. Finally, feature fusion is completed through fully connected layers, and the classification probability result is output by the Softmax output layer. The convolutional neural network model takes the floating-point sampling sequence of a single-segment cryptographic operation waveform as input, automatically extracts side-channel waveform features, completes operation classification, and outputs the cryptographic operation category label and the corresponding Softmax probability distribution of the single-segment cryptographic operation waveform.

[0026] In this embodiment of the invention, specialized and differentiated datasets are used to complete independent model training for power consumption channels and electromagnetic channels. The convolutional neural network model is independently iteratively trained using pure power consumption side-channel waveform datasets and pure electromagnetic side-channel waveform datasets, respectively, to obtain a power consumption channel cryptographic operation classification model and an electromagnetic channel cryptographic operation classification model, adapting to different channel leakage characteristics. For example... Figure 3 As shown, this is the electromagnetic side-channel leakage waveform trajectory corresponding to the execution process of the elliptic curve cryptography algorithm. This electromagnetic trajectory can clearly reflect the feature change pattern in the cryptographic operation process, providing effective data support for model feature learning.

[0027] It is understandable that the number of parameters, computational load, and activation memory usage of a fully trained convolutional neural network model is large, far exceeding the hardware capacity of a resource-constrained microcontroller, making direct deployment and operation impossible. To solve this hardware adaptation problem, this embodiment of the invention employs knowledge distillation technology to achieve lightweight model optimization, resulting in a deployment model adapted for microcontroller operation. Knowledge distillation technology uses a high-precision original large model as the teacher model and a lightweight small model as the student model. Through knowledge transfer, the small-scale student model learns the feature extraction capabilities and classification logic of the teacher model. While basically maintaining the model's classification accuracy, it significantly reduces the number of model parameters, computational load, and memory usage, achieving lightweight model iteration. Figure 4 The diagram illustrates the changes in model accuracy and loss function during model distillation training in the specific implementation of knowledge distillation technology. By observing the iterative convergence changes in training accuracy and loss value in the diagram, the feasibility and effectiveness of the knowledge distillation technology used in this invention are intuitively demonstrated, proving that the lightweight model can maintain excellent classification performance while significantly reducing resource consumption.

[0028] After obtaining the deployment model adapted to the microcontroller, an embedded neural network generation tool is used to convert the model files of each deployment model into compilable code, and then integrate the compilable code into the microcontroller projects of the first and second microcontrollers respectively.

[0029] In practical applications, after the model is lightweighted and compressed, embedded neural network deployment tools such as STM32Cube.AI can be used to automatically convert the dual-channel deployment model into standard C language model code that can be directly compiled, linked, and run by a microcontroller. The model code is then deeply integrated into the microcontroller's underlying project, completing the model's embedded adaptation and porting. In a preferred embodiment, both microcontrollers can be equipped with Cortex-M core STM32F407 microcontrollers with floating-point arithmetic units as core edge inference hardware, possessing sufficient floating-point computing power and storage resources to meet the real-time inference requirements of the deployed model. This embodiment of the invention adopts a dual-microcontroller independent division-of-labor architecture. The first microcontroller loads the deployment model corresponding to the power consumption channel, dedicated to feature inference and classification of the power consumption side channel waveform; the second microcontroller loads the deployment model corresponding to the electromagnetic channel, dedicated to intelligent inference and classification of the electromagnetic side channel waveform.

[0030] Meanwhile, to avoid problems such as inference latency fluctuations, memory overflows, and poor program stability caused by dynamic memory allocation of the microcontroller, this embodiment of the invention adopts a static memory pre-allocation mechanism in the microcontroller inference program design. It pre-allocates the feature activation memory and input / output cache memory required for model inference. There are no dynamic memory allocation and release operations during the entire inference process, which completely ensures the determinism of inference latency and the stability of memory usage of the dual-channel microcontroller, and meets the real-time and stable side-channel analysis and inference requirements of edge devices.

[0031] In step S102, the host computer establishes communication with the first microcontroller and the second microcontroller through two independent serial ports, respectively, and pushes the corresponding channel type cryptographic operation waveform data to the first microcontroller and the second microcontroller using a packet streaming transmission method with handshake mechanism.

[0032] This invention constructs a packet-splitting serial communication architecture with a handshake mechanism. The host computer is configured with two independent serial communication links, corresponding to the microcontroller devices of the power consumption channel and the electromagnetic channel, respectively. During operation, the host computer pushes the cryptographic operation waveform data of the corresponding channel to the dual microcontrollers in parallel through the two independent serial ports, realizing the synchronous transmission of data between the two channels.

[0033] In this embodiment of the invention, a packet-splitting serial communication protocol with handshake verification, error detection, and timing control capabilities was custom-designed to address the requirements of edge device transmission stability, data integrity, and timing control. For example... Figure 5 As shown, the protocol data frame structure consists of a start flag, command word, sequence number, payload length, payload data, and a cyclic redundancy check (CRC) field. The start flag precisely identifies the beginning of a single frame, preventing frame misalignment and data packet merging. The command word is the core control field of the protocol, accurately distinguishing between various frame types such as handshake frames, session start frames, segment start frames, streaming data frames, segment end frames, response frames, and error frames, enabling state management at different communication stages. The sequence number matches requests sent by the host computer with responses uploaded by the microcontroller, ensuring the timing correspondence of multi-segment waveform data transmission. The payload length precisely identifies the number of bytes of valid data in the current frame, facilitating data parsing and length verification at the receiving end. The CRC field detects data loss, errors, interference, and distortion during data transmission in real time, ensuring transmission reliability in complex edge scenarios. Furthermore, to accurately reflect the microcontroller's operating status, a status code field is added at the beginning of the payload data in the response frame, indicating normal reception, verification error, inference anomaly, device idle, and device busy statuses, allowing the host computer to monitor the operating status of the dual-channel microcontroller in real time.

[0034] Based on the aforementioned packet-based serial communication protocol, this embodiment of the invention divides the complete dual-channel data transmission and inference interaction process into three time-series stages: the handshake stage, the data transmission stage, and the aggregation stage. Each stage is logically independent, sequentially progressive, and state controllable. The handshake stage is the communication initialization stage. The host computer actively sends a handshake frame to the dual-channel microcontrollers to wake up the microcontrollers and activate the link. After receiving the handshake command, the microcontrollers return an acknowledgment frame carrying communication parameters such as their baud rate, buffer size, and model inference latency, completing the matching of communication parameters between the host and microcontrollers and link calibration, thus establishing a reliable data transmission channel. The data transmission stage is the core working stage. The host computer first sends a session start frame carrying the total number of cryptographic waveform segments and the algorithm type, informing the microcontrollers of the total task and computation identification type for this session. Subsequently, for each segment of cryptographic waveform data, it sequentially sends a segment start frame, multiple streaming data frames, and a segment end frame, realizing the packet-based streaming transmission of large-size waveform data and avoiding packet loss and stuttering problems caused by excessive data volume in a single frame. The microcontroller continuously receives and parses streaming data frames, writing waveform sampling points into the neural network input tensor in real time. Upon receiving the end frame of the corresponding segment, it determines that the reception of a single segment of cryptographic waveform data is complete, automatically triggers local neural network inference operations, and immediately returns the inference result containing cryptographic operation category labels and the corresponding Softmax probability distribution. The summary phase is the session termination phase. After all cryptographic waveform data transmission and inference are completed, the host computer sends a session termination frame to the microcontroller. Upon receiving this frame, the microcontroller can return an acknowledgment, automatically reset the communication state and model input tensor, release temporary resources, and await the next round of session tasks, thus realizing multi-batch, continuous side-channel analysis and inference.

[0035] In step S103, the first microcontroller and the second microcontroller respectively use the deployed neural network model to perform neural network inference on the cryptographic operation waveform data, output the inference result containing the cryptographic operation category label and the corresponding Softmax probability distribution, and send the inference result back to the host computer.

[0036] In this embodiment of the invention, the microcontroller abandons the traditional processing method of first caching the entire waveform before inputting it into the model. Instead, it adopts an innovative mechanism of streaming reception and in-situ writing to the neural network input tensor. This eliminates the need for a large-capacity cache to store the complete waveform data, greatly alleviating the pressure on the microcontroller's limited memory resources. After receiving the single-segment cryptographic operation waveform data, the neural network inference is immediately triggered, and the obtained cryptographic operation classification labels and Softmax probability distribution are encapsulated into a response frame and sent back to the host computer to complete data aggregation, providing reliable underlying data support for subsequent multi-evidence fusion.

[0037] Specifically, during the initialization phase, the microcontroller pre-allocates a fixed-size activation memory space for the neural network inference process using static memory allocation, without performing dynamic memory allocation or deallocation operations, ensuring stable inference latency and controllable memory usage. Simultaneously, it directly obtains the dedicated memory pointer of the neural network input tensor, locking the physical memory address of the model input region, providing the address basis for subsequent in-situ writing of streaming data. The microcontroller's overall firmware operates based on a state machine mechanism, relying on protocol command words to precisely trigger device state transitions, responding only to commands with valid timing and format, and directly filtering out-of-order frames, erroneous frames, and abnormal commands, significantly improving device stability and anti-interference capabilities. To adapt to high-speed, continuous waveform streaming scenarios, a composite data reception mechanism combining DMA (Direct Memory Access) circular buffering with serial port idle interrupts can also be used. DMA hardware automatically completes data transfer, reducing CPU resource consumption. Upon triggering a serial port idle interrupt, frame structure parsing and CRC (Cyclic Redundancy Check) transmission integrity verification are performed immediately, ensuring the accuracy of each frame of data.

[0038] During the reception of streaming data frames, the microcontroller, based on the neural network input tensor pointer acquired during the initialization phase, writes the waveform sampling points within the current streaming data frame payload to the corresponding offset positions of the neural network input tensor in situ through precise memory copying. This process is repeated frame-by-frame to continuously fill all sampling points of a single segment of the cryptographic operation waveform, eliminating the need for additional temporary buffers to store the entire waveform data and fundamentally reducing microcontroller memory resource overhead. Once the segment's end frame is received and all single-segment cryptographic operation waveform data is filled into the input tensor, the microcontroller immediately triggers and executes lightweight neural network inference operations, outputting the cryptographic operation classification label and Softmax probability distribution result corresponding to the current waveform. This achieves a low-latency, low-memory streaming inference mechanism that "receives, fills, and infers immediately after reception," perfectly adapting to edge deployment scenarios with resource-constrained microcontrollers.

[0039] In step S104, the host computer, based on the Dempster-Shafer evidence theory, uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fusion probability distributions corresponding to each cryptographic operation waveform data.

[0040] This invention employs a decision-level dual-channel evidence fusion strategy, fully leveraging the independence and complementarity of leakage characteristics in the power consumption channel and the electromagnetic channel. This addresses the technical problems of low classification confidence and unstable judgment results caused by environmental noise and acquisition interference in a single channel, significantly improving the accuracy and robustness of cryptographic classification. A schematic diagram of the Dempster-Shafer evidence fusion principle is shown below. Figure 6 As shown, a recognition framework adapted for public-key cryptographic operations is first constructed. This identification framework includes two mutually exclusive and complete categories of cryptographic operations. , It covers the core computational behaviors of mainstream public-key cryptosystems. , Different cryptographic operations can be adaptively applied based on the type of public-key cryptography algorithm being analyzed: for example, for elliptic curve cryptography algorithms, Addition operation at corresponding points Corresponding to doubling operations; for the RSA algorithm, Corresponding to modular multiplication, This corresponds to the modulo square operation; for modulo-based key encapsulation mechanisms, and These correspond to the operational characteristics when the message bits are 0 or 1, respectively. This enables a generalized side-channel analysis for various public-key cryptographic algorithms.

[0041] Furthermore, the power consumption channel cryptographic operation classification model is targeted at... , The softmax probability distribution of the output of the two types of cryptographic operations is denoted as follows: The electromagnetic channel cryptographic operation classification model is targeted at , The softmax probability distribution of the output of the two types of cryptographic operations is denoted as follows: Both probability distributions are output by a convergent normalized neural network, satisfying the non-negativity and normalization constraints of the basic probability assignment in Dempster-Shafer evidence theory. No additional normalization processing is required, and they can be directly used as two sets of independent discriminative evidence in the fusion calculation.

[0042] This invention calculates the conflict amount of the two basic probability assignments based on a preset Dempster combination rule. This is used to quantify the degree of difference between the inference results of the power channel and the electromagnetic channel, and to accurately determine the consistency of evidence. The mathematical expression is as follows:

[0043] in, Indicates the number of conflicts; , These represent the first microcontroller targeting... , Softmax probability distributions of the outputs of two types of cryptographic operations; , These respectively indicate that the second microcontroller is targeting , The softmax probability distribution of the output of two types of cryptographic operations. The closer the value is to 1, the greater the difference between the two evidence results, indicating a high degree of evidence conflict. Conventional fusion calculation is prone to problems such as result distortion and judgment failure.

[0044] After solving for the collision quantity, this embodiment of the invention eliminates the probability bias caused by evidence collisions based on a normalized collision factor, and obtains the fused global basic probability assignment, thus completing dual-channel evidence fusion. Specifically, adaptive evidence fusion is performed based on the collision quantity to obtain the fusion probability distribution of the cryptographic operation category corresponding to the single-segment cryptographic operation waveform data, as shown in the following mathematical expression:

[0045]

[0046] in, Cryptographic operation category representing single-segment cryptographic waveform data The corresponding fusion probability distribution; Indicates the category of cryptographic operations The corresponding fusion probability distribution; This represents the normalized conflict factor, ensuring that the probability distribution after fusion satisfies the normalization condition, and that the probability result is legal and valid.

[0047] To address the shortcomings of traditional Dempster fusion rules in high-collision scenarios, this invention incorporates a high-collision adaptive fault-tolerant mechanism. When the collision amount... When the probability approaches 1, the two channels are considered to be in a state of complete conflict. At this point, the traditional fusion calculation logic is abandoned, and the single-channel result with the higher Softmax probability and stronger discrimination confidence among the two single-channel inference results is adaptively selected as the final discrimination criterion for the current segment's cryptographic operation. This mechanism effectively avoids the fusion failure problem in high-noise and strong interference scenarios, and greatly improves the robustness and environmental adaptability of the overall analysis method.

[0048] Step S105: Determine the target cryptographic operation category of each cryptographic operation waveform data based on the fusion probability distribution, and concatenate them in chronological order to obtain a cryptographic operation discrimination result sequence. Recover the key of the public key cryptographic algorithm to be analyzed based on the operation discrimination result sequence.

[0049] Specifically, for each segment of collected and analyzed cryptographic waveform data, the cryptographic operation category with the highest fusion probability distribution is selected as the target cryptographic operation category. Specifically, this is achieved by comparing the fusion probability distributions. and The numerical value is used to select the cryptographic operation category with a larger fusion probability distribution as the target cryptographic operation category corresponding to that waveform segment. After determining the target cryptographic operation category for each individual cryptographic operation waveform segment by segment, all target cryptographic operation categories are arranged strictly according to the chronological order of waveform acquisition time to form a complete, time-corresponding cryptographic operation discrimination result sequence.

[0050] Based on the mapping relationship between the inherent operational logic of various public-key cryptography algorithms and key bits, the discrimination result sequence is parsed, and the key bit information is recovered bit by bit to finally complete the recovery of the complete key. The specific adaptation logic is as follows: For scalar multiplication operations in elliptic curve cryptography, the private key bits are parsed and recovered bit by bit based on the temporal combination pattern of point addition and doubling operations; for modular exponentiation operations in RSA cryptography, the key is recovered by matching key bit characteristics through the alternation pattern of modular multiplication and modular squaring operations; for modular lattice-based key encapsulation mechanisms, a single discrimination result can directly correspond to the 0 or 1 values ​​of the decrypted message bits, quickly completing message key parsing.

[0051] To fully verify the engineering practicality, stability, and accuracy of the public-key cryptography AI side-channel analysis method proposed in this invention, a complete dual-channel embedded side-channel analysis experimental platform was built to perform multiple sets of comparative experiments and performance tests. The experimental hardware uses two STM32F407 microcontrollers as power consumption and electromagnetic channel edge inference terminals, respectively. Streaming data transmission between the host computer and the microcontrollers is achieved via a 921600bps serial port baud rate. The entire process realizes real-time waveform data transmission, in-situ inference, and instant result feedback. The entire system operates stably without packet loss, lag, or inference anomalies, and can adapt to automated analysis tasks involving multiple batches of cryptographic waveforms. Specifically, three mainstream public-key cryptography algorithms—ML-KEM, ECC, and RSA—were selected as test objects, and multiple sets of comparative experiments were conducted. The experiment compared the traditional PCA&K-Means, SVD, and GSVD classical side-channel analysis methods with the CNN-DS dual-channel fusion method of this invention. Using the number of waveforms as a variable, the accuracy rates of power consumption channel identification, electromagnetic channel identification, and dual-channel fusion for each method were statistically analyzed. This comprehensively and quantitatively verified the optimization effect of the single-channel feature extraction capability and dual-channel evidence fusion of this invention. The specific experimental data comparison results are shown in Table 1 below: Table 1 (Comparative Experiment Results)

[0052] Comparative experimental results show that, compared with traditional side-channel analysis schemes that rely on a single power consumption channel or a single electromagnetic channel, the dual-channel evidence fusion scheme adopted in this invention has significant comprehensive performance advantages. Traditional single-channel analysis methods are highly susceptible to environmental noise and acquisition interference, resulting in low recognition accuracy and poor result stability. In contrast, this invention, through a dual-channel independent reasoning combined with evidence fusion architecture, fully leverages the complementarity of leakage characteristics of the power consumption channel and the electromagnetic channel, effectively compensating for the shortcomings of single-channel recognition. It achieves significant improvements in power consumption recognition accuracy, electromagnetic recognition accuracy, and dual-channel fusion accuracy, effectively solving the technical pain points of easy confusion and high false positive rate in cryptographic operation classification under low signal-to-noise ratio and complex interference scenarios, and significantly improving the accuracy and robustness of public-key cryptography AI side-channel analysis.

[0053] like Figure 7 The figure shows a comparison curve of the classification confidence of the RSA algorithm. This figure is used to visually compare the classification confidence trends of single-channel and dual-channel fusion methods under low signal-to-noise ratio and electromagnetic noise conditions. Figure 7 The following confidence curves were plotted: (a) for single electromagnetic channel inference, (b) for single power consumption channel inference, and (c) for the dual-channel evidence fusion inference of this invention. Combined with... Figure 7 Further analysis reveals that traditional single-channel models, under complex electromagnetic interference and low signal-to-noise ratio conditions in real-world engineering scenarios, consistently exhibit classification confidence curves hovering around the 0.5 critical decision threshold. This results in highly uncertain model outputs, making them prone to confusion between adjacent cryptographic operation categories and misjudgments of consecutive frames. Consequently, they struggle to stably and accurately identify cryptographic operations, ultimately leading to key parsing failures. Furthermore, classic algorithms such as PCA&K-Means, SVD, and GSVD rely on manual feature extraction, limiting their ability to capture weak side-channel leakage features. In single-waveform sample inference scenarios, their recognition accuracy significantly decreases, exhibiting poor adaptability and robustness. In contrast, the CNN-DS dual-channel fusion method proposed in this invention can accurately capture subtle leakage features across different dimensions through independent dual-channel neural networks. Leveraging the Dempster-Shafer evidence fusion mechanism, it achieves multi-source evidence complementarity and conflict correction. The fused classification confidence is significantly enhanced, effectively escaping the 0.5 fuzzy decision interval, resulting in stable and reliable judgment results. The performance variation pattern shown in the figure fully demonstrates that the dual-channel fusion architecture can effectively suppress environmental noise interference, make up for the technical shortcomings of insufficient feature leakage and susceptibility to interference in single-channel architecture, and exhibit excellent recognition stability and generalization ability under various public key cryptography algorithm datasets. It can meet the high-precision and high-reliability side-channel key analysis and recovery requirements in complex engineering scenarios.

[0054] This invention provides a public-key cryptography AI side-channel analysis method. It constructs neural network models for cryptographic operation classification, targeting power consumption channels and electromagnetic channels respectively. Using neural network compression technology, each neural network model is independently deployed to a first microcontroller and a second microcontroller. A host computer establishes communication with the first and second microcontrollers via two independent serial ports, employing a packet-splitting streaming transmission method with a handshake mechanism to push cryptographic operation waveform data corresponding to the channel type to the first and second microcontrollers respectively. The first and second microcontrollers then utilize their deployed neural network models to execute neural operations on the cryptographic operation waveform data. The network inference process outputs an inference result containing cryptographic operation category labels and corresponding Softmax probability distributions, and sends the inference result back to the host computer. Based on Dempster-Shafer evidence theory, the host computer uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fused probability distribution corresponding to each cryptographic operation waveform data. Based on the fused probability distribution, the target cryptographic operation category of each cryptographic operation waveform data is determined, and the results are concatenated in chronological order to obtain a cryptographic operation discrimination result sequence. The key of the public-key cryptographic algorithm to be analyzed is recovered based on the operation discrimination result sequence. Compared with the prior art, this invention has at least the following beneficial effects: First, it achieves lightweight edge deployment of multi-channel intelligent side-channel analysis: This invention uses model compression and knowledge distillation techniques to lightweight the dual-channel convolutional neural network. Combined with a single-chip microcomputer static memory deployment scheme, it pushes the deep learning side-channel analysis capability, which originally relied on PCs and servers, down to resource-constrained single-chip microcomputer edge devices. This eliminates the dependence on high-performance computing platforms and can be adapted to lightweight application scenarios such as on-site evaluation and portable testing, thus expanding the scope of application of the technology.

[0055] Second, the invention features a unique streaming in-situ tensor writing mechanism that significantly reduces memory overhead for edge devices: This invention abandons the traditional method of buffering entire waveform segments and instead receives and writes the neural network input tensor in-situ at offsets, eliminating the need for additional waveform buffer arrays throughout the process. This completely solves the core problem of limited microcontroller memory resources, ensuring stable and reliable inference at the edge and eliminating the risk of memory overflow.

[0056] Third, a reliable dual-machine collaborative streaming communication mechanism is constructed to ensure data transmission accuracy: This invention designs a dedicated serial port streaming communication protocol with handshake, frame verification, and status reset, and uses DMA circular buffer and serial port idle interrupt to achieve data reception. This effectively solves the problems of packet loss, error, and poor synchronization in traditional serial port transmission, and realizes stable and synchronous transmission of large batches of side channel waveform data between the host computer and the dual-channel microcontroller, providing data assurance for accurate collaborative inference.

[0057] Fourth, the robustness of the analysis and the accuracy of key recovery are improved by Dempster-Shafer evidence fusion: This invention integrates power consumption and electromagnetic dual-channel side-channel evidence, utilizing the complementary advantages of channel characteristics to compensate for the shortcomings of single-channel analysis, such as susceptibility to noise interference and low discrimination accuracy. Simultaneously, an adaptive decision-making mechanism for evidence conflict is added to effectively avoid misjudgments caused by evidence conflicts. The measured key recovery accuracy can reach over 98%, significantly better than traditional single-channel analysis schemes, with stronger anti-interference capabilities and generalization.

[0058] Fifth, a complete and lightweight edge side-channel analysis closed-loop architecture is formed: This invention relies on the division of labor architecture of dual-channel single-chip microcomputer collaborative reasoning, host computer evidence fusion and key parsing to realize the lightweight implementation of the entire process of waveform transmission, edge reasoning, multi-source fusion and key recovery. The overall architecture has low overhead, good real-time performance and convenient deployment, providing a feasible engineering solution for the physical security assessment of portable cryptographic devices.

[0059] Furthermore, as Figure 1 In specific implementation, embodiments of the present invention provide a public-key cryptography artificial intelligence side-channel analysis system, such as... Figure 8 As shown, the system may include: a host computer 810, a first microcontroller 820, and a second microcontroller 830.

[0060] The first microcontroller 820 and the second microcontroller 830 are each equipped with a neural network model that has been processed using neural network compression technology and classifies cryptographic operations for power consumption channels and electromagnetic channels. The host computer 810 establishes communication with the first microcontroller 820 and the second microcontroller 830 through two independent serial ports, using a packet-segmented streaming transmission method with a handshake mechanism, and pushes cryptographic operation waveform data of the corresponding channel type to the first microcontroller 820 and the second microcontroller 830 respectively. The first microcontroller 820 and the second microcontroller 830 utilize their respective deployed neural network models to perform neural network inference on the cryptographic operation waveform data, and output data containing cryptographic operation categories. The system identifies the labels and their corresponding Softmax probability distributions, and sends the inference results back to the host computer 810. The host computer 810 is also used to perform dual-channel evidence fusion based on Dempster-Shafer evidence theory, using the Softmax probability distributions sent back by the first microcontroller 820 and the second microcontroller 830 as basic probability assignments to obtain the fused probability distributions corresponding to each cryptographic operation waveform data. Based on the fused probability distributions, the system determines the target cryptographic operation category of each cryptographic operation waveform data, and concatenates them in chronological order to obtain a cryptographic operation discrimination result sequence. Based on the operation discrimination result sequence, the system recovers the key of the public-key cryptographic algorithm to be analyzed.

[0061] It should be noted that other corresponding descriptions of the functional modules involved in the public-key cryptography artificial intelligence side-channel analysis system provided in this embodiment of the invention can be found in [reference]. Figure 1 The corresponding description of the method shown will not be repeated here.

[0062] Those skilled in the art will clearly understand that the specific working process of the systems, devices, modules and units described above can be referred to the corresponding process in the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0063] Furthermore, the functional units in the various embodiments of the present invention can be physically independent of each other, or two or more functional units can be integrated together, or all functional units can be integrated into one processing unit. The integrated functional units described above can be implemented in hardware, or in software or firmware.

[0064] Those skilled in the art will understand that if the integrated functional unit is implemented in software and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or all or part of it, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computing device (e.g., a personal computer, server, or network device) to execute all or part of the steps of the methods described in the embodiments of the present invention when running the instructions. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] Alternatively, all or part of the steps of the foregoing method embodiments can be implemented by hardware (such as a computing device, personal computer, server, or network device) related to program instructions. The program instructions can be stored in a computer-readable storage medium. When the program instructions are executed by the processor of the computing device, the computing device executes all or part of the steps of the methods described in the various embodiments of the present invention.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of the present invention, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to depart from the protection scope of the present invention.

Claims

1. A public-key cryptography artificial intelligence side-channel analysis method, characterized in that, The method includes: Neural network models for cryptographic operation classification are constructed for power consumption channels and electromagnetic channels respectively, and neural network compression technology is used to independently deploy each neural network model to the first microcontroller and the second microcontroller. The host computer establishes communication with the first microcontroller and the second microcontroller through two independent serial ports respectively, and uses a packet streaming transmission method with handshake mechanism to push cryptographic operation waveform data of the corresponding channel type to the first microcontroller and the second microcontroller respectively; The first microcontroller and the second microcontroller respectively use the deployed neural network model to perform neural network inference on the cryptographic operation waveform data, output the inference result containing the cryptographic operation category label and the corresponding Softmax probability distribution, and send the inference result back to the host computer; The host computer, based on the Dempster-Shafer evidence theory, uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fusion probability distributions corresponding to each of the cryptographic operation waveform data. The target cryptographic operation category of each cryptographic operation waveform data is determined based on the fusion probability distribution, and the cryptographic operation discrimination result sequence is obtained by splicing them in chronological order. The key of the public key cryptographic algorithm to be analyzed is recovered based on the operation discrimination result sequence.

2. The method according to claim 1, characterized in that, The construction of neural network models for cryptographic operation classification, specifically for power consumption channels and electromagnetic channels, includes: Two independent convolutional neural network models are constructed for the power consumption channel and the electromagnetic channel, respectively. Each convolutional neural network model includes several convolutional layers, batch normalization layers, pooling layers, global average pooling layers, several fully connected layers, and a Softmax output layer. The convolutional neural network model takes the floating-point sampling sequence of a single-segment cryptographic operation waveform as input, automatically extracts side-channel waveform features, completes operation classification, and outputs the cryptographic operation category label and the corresponding Softmax probability distribution of the single-segment cryptographic operation waveform. Each convolutional neural network model is trained independently using dedicated training datasets for the power consumption channel and the electromagnetic channel, respectively, to obtain a power consumption channel cryptographic operation classification model and an electromagnetic channel cryptographic operation classification model.

3. The method according to claim 1, characterized in that, The method employs neural network compression technology to independently deploy each neural network model to the first microcontroller and the second microcontroller, including: Knowledge distillation technology is used to lightweight each neural network model to obtain a deployment model adapted for microcontroller operation; An embedded neural network generation tool is used to convert the model files of each deployment model into compilable code, and the compilable code is integrated into the microcontroller projects of the first microcontroller and the second microcontroller, respectively. The first microcontroller deploys the deployment model corresponding to the power consumption channel; the second microcontroller deploys the deployment model corresponding to the electromagnetic channel.

4. The method according to claim 1, characterized in that, The packet streaming transmission method with handshake mechanism is implemented using a preset serial communication protocol, including a handshake phase, a data transmission phase, and a summarization phase. During the handshake phase, the host computer sends a handshake frame to the microcontroller, and the microcontroller replies with an acknowledgment frame carrying communication parameters, thus completing the communication link initialization. During the data transmission phase, the host computer first sends a session start frame to carry the total number of cryptographic waveform segments and the algorithm type. Then, for each cryptographic waveform data segment, it sequentially sends a segment start frame, a streaming data frame carrying the cryptographic waveform data, and a segment end frame to the microcontroller. When the microcontroller receives the streaming data frame, it synchronously writes the waveform sampling points contained in the streaming data frame into the neural network input tensor. After receiving the segment end frame, it triggers the neural network inference and sends the inference result back to the host computer. During the aggregation phase, the host computer sends a session end frame to the microcontroller, and the microcontroller resets its own state after receiving the session end frame.

5. The method according to claim 4, characterized in that, In the preset serial communication protocol, the frame structure includes a start flag, command word, sequence number, payload length, payload data, and cyclic redundancy check field. The start flag is used to identify the starting position of a single frame of data. The command word is used to distinguish frame types; the frame types include at least one of handshake frames, session start frames, segment start frames, streaming data frames, segment end frames, response frames, and error frames. The sequence number is used to establish the association between request frames and response frames; The payload length is used to identify the number of bytes of payload data; The cyclic redundancy check field is used to identify data transmission errors; The response frame is used to add a status code field at the beginning of the payload data to identify the data interaction status.

6. The method according to claim 4, characterized in that, The first microcontroller and the second microcontroller respectively utilize the deployed neural network model to perform neural network inference on the cryptographic waveform data, including: During initialization, the microcontroller pre-allocates activation memory for the neural network inference using a static memory allocation method and obtains the neural network input tensor pointer. The microcontroller firmware runs based on a state machine mechanism. It triggers the transition of working state through command words and responds to valid command instructions. At the same time, it uses DMA circular buffer combined with serial port idle interrupt to receive data. After the interrupt is triggered, it parses the frame structure and completes CRC transmission verification. When the microcontroller receives a streaming data frame, it copies the waveform sampling points within the streaming data frame payload to the corresponding offset positions of the neural network input tensor in situ via memory copying, based on the acquired neural network input tensor pointer, and completes the filling of all waveform sampling points in the single-segment cryptographic operation waveform data frame by frame; after all the single-segment cryptographic operation waveform data has been received and filled, the neural network inference is executed.

7. The method according to claim 1, characterized in that, The host computer, based on Dempster-Shafer evidence theory, uses the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments, performs dual-channel evidence fusion, and obtains the fused probability distributions corresponding to each of the cryptographic waveform data, including: Establish an identification framework that includes two categories of cryptographic operations. The two Softmax probability distributions output by the first microcontroller and the second microcontroller are used as the two basic probability values ​​corresponding to the Dempster-Shafer evidence theory. The conflict amount between the two basic probability assignments is calculated using the preset Dempster combination rule, and the mathematical expression is as follows: in, Indicates the number of conflicts; , These respectively indicate that the first microcontroller is targeting , Softmax probability distributions of the outputs of two types of cryptographic operations; , These respectively indicate that the second microcontroller is targeting , Softmax probability distributions of the outputs of two types of cryptographic operations; Adaptive evidence fusion is performed based on the collision amount to obtain the fusion probability distribution of the cryptographic operation category corresponding to the single-segment cryptographic operation waveform data. The mathematical expression is as follows: in, Cryptographic operation category representing single-segment cryptographic waveform data The corresponding fusion probability distribution; Indicates the category of cryptographic operations The corresponding fusion probability distribution.

8. The method according to claim 1, characterized in that, The step of determining the target cryptographic operation category of each cryptographic operation waveform data based on the fusion probability distribution, concatenating them in chronological order to obtain a cryptographic operation discrimination result sequence, and recovering the key of the public-key cryptography algorithm to be analyzed based on the operation discrimination result sequence includes: For single-segment cryptographic operation waveform data, the cryptographic operation category with the largest fusion probability distribution is selected as the target cryptographic operation category; Arrange the target cryptographic operation categories corresponding to all the single-segment cryptographic operation waveform data in the order of waveform acquisition time to obtain a cryptographic operation discrimination result sequence; Based on the mapping relationship between the operation type of the public-key cryptography algorithm to be analyzed and the key bits, the sequence of cryptographic operation discrimination results is parsed bit by bit to obtain the key of the public-key cryptography algorithm.

9. The method according to any one of claims 1 to 8, characterized in that, The public-key cryptography algorithm includes at least one of elliptic curve cryptography, RSA algorithm, and lattice-based key encapsulation mechanism.

10. A public-key cryptography artificial intelligence side-channel analysis system, characterized in that, The system is used to execute the public-key cryptography artificial intelligence side-channel analysis method according to any one of claims 1 to 9, the system comprising a host computer, a first microcontroller, and a second microcontroller; The first microcontroller and the second microcontroller are respectively deployed with neural network models that have been processed by neural network compression technology and perform cryptographic operations to classify power consumption channels and electromagnetic channels; The host computer is used to establish communication with the first microcontroller and the second microcontroller through two independent serial ports respectively, and adopts a packet streaming transmission method with handshake mechanism to push cryptographic operation waveform data of corresponding channel type to the first microcontroller and the second microcontroller respectively. The first microcontroller and the second microcontroller are respectively used to perform neural network inference on the cryptographic operation waveform data using the deployed neural network model, output the inference result containing the cryptographic operation category label and the corresponding Softmax probability distribution, and send the inference result back to the host computer; The host computer is also used to perform dual-channel evidence fusion based on Dempster-Shafer evidence theory, using the Softmax probability distributions returned by the first and second microcontrollers as basic probability assignments to obtain the fusion probability distributions corresponding to each of the cryptographic operation waveform data; determine the target cryptographic operation category of each of the cryptographic operation waveform data according to the fusion probability distributions, and concatenate them in chronological order to obtain a cryptographic operation discrimination result sequence; and recover the key of the public-key cryptography algorithm to be analyzed based on the operation discrimination result sequence.

Images