Hardware Trojan detection method and system based on multi-modal side channel fusion

Abstract

The application relates to the technical field of hardware Trojan detection, and discloses a hardware Trojan detection method and system based on multi-modal side channel fusion, which comprises a signal sensing and fusion module, a detection strategy matching and execution module, and an evaluation and optimization module. Static and dynamic multi-modal side channel data of a to-be-detected chip are collected, preprocessed and feature-fused to obtain multi-modal fusion feature data of the chip, which is matched with a pre-constructed standard feature template library to screen out a target detection algorithm, finally completing Trojan judgment and positioning, realizing comprehensive utilization of multi-modal side channel features and intelligent algorithm matching, more comprehensively representing chip behavior, and improving the detection rate and detection accuracy of hardware Trojans. Multi-dimensional side channel data containing static and dynamic characteristics are collected, adaptive weight distribution and fusion are realized by using an attention mechanism, and deep processing and key feature enhancement of multi-source heterogeneous data are realized.

Description

Technical Field

[0001] This invention relates to the field of hardware Trojan detection technology, specifically to a hardware Trojan detection method and system based on multimodal side-channel fusion. Background Technology

[0002] Hardware Trojans are malicious circuit modules that are intentionally implanted during the design or manufacturing stage of integrated circuits. They are characterized by their strong concealment and high destructiveness. Once triggered, they can lead to information leakage, system function tampering, or even physical damage. Common implantation methods include third-party IP cores or manufacturing processes. Defense strategies cover logic testing, trusted design, and more.

[0003] Currently, in the field of integrated circuit hardware Trojan detection, existing methods usually rely on a single or a few types of physical side-channel information for analysis, failing to effectively integrate the static and dynamic multi-modal side-channel characteristics of the chip. This results in an incomplete characterization of chip behavior and makes it difficult to capture the faint traces of hidden or small-sized Trojans.

[0004] Therefore, a hardware Trojan detection method and system based on multimodal side-channel fusion is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a hardware Trojan detection method and system based on multimodal side-channel fusion, which solves the problem mentioned in the background technology that the characterization of chip behavior is incomplete and it is difficult to capture the weak traces of some hidden or small-sized Trojans.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a hardware Trojan detection method and system based on multimodal side-channel fusion, wherein the method includes the following steps: S1. Collect static multimodal side-channel data and dynamic multimodal side-channel data of the integrated circuit chip under test; S2. Preprocess and align the static multimodal side channel data and dynamic multimodal side channel data with the data acquisition frequency to generate standardized static multimodal side channel feature data and standardized dynamic multimodal side channel feature data. S3. Perform multi-mode side-channel feature fusion processing based on the standardized static multi-mode side-channel feature data and the standardized dynamic multi-mode side-channel feature data to generate chip multi-mode fusion feature data; S4. Construct a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms; S5. Match the chip multimodal fusion feature data with the standard chip multimodal fusion feature template library to filter out the target hardware Trojan detection algorithm type; S6. Based on the target hardware Trojan detection algorithm type and the chip multimodal fusion feature data, perform hardware Trojan existence determination and location analysis on the integrated circuit chip under test, and generate a primary hardware Trojan detection report. S7. Perform result credibility calibration and optimization on the primary hardware Trojan detection report to generate the final hardware Trojan detection report.

[0007] Preferably, the acquisition of static multimode side-channel data and dynamic multimode side-channel data of the integrated circuit chip under test in S1 includes the following steps: S11. Collect multi-mode side-channel data of the integrated circuit chip under test at the unpowered static operating point using an integrated circuit static parameter analyzer, and generate static multi-mode side-channel data, wherein the static multi-mode side-channel data includes the chip's static current data, static electromagnetic radiation data, static thermal distribution data, and chip die microscopic image data. S12. Collect multi-modal side-channel data of the integrated circuit chip under test under dynamic operation with preset test vectors loaded through the integrated circuit dynamic test platform, and generate dynamic multi-modal side-channel data. The dynamic multi-modal side-channel data includes dynamic power consumption trajectory data, transient electromagnetic radiation data, runtime sequence data and dynamic thermal imaging data of the chip surface.

[0008] Preferably, the preprocessing and data acquisition frequency alignment process in S2 includes the following steps: S21. Perform data cleaning on the static multimodal side channel data and the dynamic multimodal side channel data respectively, remove outliers and noise, and perform normalization processing to generate preprocessed static multimodal side channel data and preprocessed dynamic multimodal side channel data. S22. Identify the data acquisition frequencies of the preprocessed static multimodal side channel data and the preprocessed dynamic multimodal side channel data. If the frequencies are inconsistent, use an interpolation algorithm to adjust the data acquisition frequency of the preprocessed static multimodal side channel data to be consistent with the data acquisition frequency of the preprocessed dynamic multimodal side channel data, and generate frequency-aligned static multimodal side channel data. S23. Perform multi-scale feature extraction on the frequency-aligned static multimodal side-channel data and the preprocessed dynamic multimodal side-channel data respectively, extract key features in the time domain, frequency domain, and time-frequency domain, and generate standardized static multimodal side-channel feature data and standardized dynamic multimodal side-channel feature data.

[0009] Preferably, the multimodal side-channel feature fusion processing in S3 includes the following steps: S31. Concatenate the standardized static multimodal side-channel feature data with the standardized dynamic multimodal side-channel feature data to construct the original multimodal feature vector. S32. An attention mechanism is used to adaptively assign weights to the features of different modalities in the original multimodal feature vector, thereby enhancing the modal features that are highly correlated with the behavior of the hardware Trojan and suppressing irrelevant and noisy modal features. S33. Perform dimensionality reduction and fusion on the multimodal features after weight allocation to generate chip multimodal fusion feature data.

[0010] Preferably, the step S4, which involves constructing a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms, includes the following steps: S41. Establish a multimodal fusion feature template library matrix for standard chips corresponding to different hardware Trojan detection algorithms. , ,in Indicates the first The standard chip multimodal fusion feature vector corresponding to the various hardware Trojan detection algorithm types This represents the maximum number of hardware Trojan detection algorithm types. S42. The hardware Trojan detection algorithm model types include classification algorithms based on support vector machines, anomaly detection algorithms based on deep neural networks, template matching algorithms based on side-channel power consumption analysis, and detection algorithms based on timing violation analysis. The standard chip multimodal fusion feature vector... This represents the standard feature template obtained by using the first detection algorithm to perform multimodal side-channel data acquisition, preprocessing, feature extraction, and fusion on a known Trojan-free gold chip.

[0011] Preferably, step S5 includes the following steps: S51. Calculate the matching degree between the chip multimodal fusion feature data and each standard feature vector in the standard chip multimodal fusion feature template library matrix; S52. The improved Osprey optimization algorithm is used to search for the hardware Trojan detection algorithm type corresponding to the standard feature vector with the highest matching degree with the chip multimodal fusion feature data, and this algorithm type is determined as the target hardware Trojan detection algorithm type. The specific search steps are as follows: S521. Initialize the parameters of the improved Osprey optimization algorithm, including the Osprey population size and the maximum number of iterations, and randomly initialize the position of the Osprey population in the algorithm search space, which corresponds to different candidate detection algorithm types. S522, Exploration Phase: Simulate the behavior of ospreys recognizing and exploring prey. Individual ospreys randomly explore different detection algorithm types in the search space and calculate their matching degree with the chip's multimodal fusion feature data as a fitness value. Update the osprey's position based on the fitness value. S523. During the development phase, the behavior of ospreys capturing and consuming prey is simulated. Individual ospreys conduct a detailed search in the vicinity of their current location to find a detection algorithm type with better matching degree and update their position accordingly. S524. Iteratively execute the exploration and development phase until the maximum number of iterations is met, and output the hardware Trojan detection algorithm type corresponding to the Osprey position with the highest fitness value, i.e. the best matching degree, as the target hardware Trojan detection algorithm type.

[0012] Preferably, the hardware Trojan existence determination and location analysis in S6 includes the following steps: S61. Combine the chip multimodal fusion feature data with the target hardware Trojan detection algorithm type to construct a hardware Trojan detection task package; S62. Based on the target hardware Trojan detection algorithm type, call the corresponding detection algorithm model; S63. Input the chip's multimodal fusion feature data into the called detection algorithm model, perform inference calculations, and output the preliminary Boolean determination result of the Trojan's existence and the location information of the circuit module where the Trojan is located. S64. Integrate the results of the Trojan existence determination, location information, the type of algorithm used, and a summary of the feature data to generate a basic hardware Trojan detection report.

[0013] Preferably, the result reliability calibration and optimization process in S7 includes the following steps: S71. Perform a reliability analysis on the primary hardware Trojan detection report, including calculating the confidence score of the detection results and analyzing the contribution of different modal features to the detection results. S72. Based on contribution analysis, re-acquire and verify the original side channel data corresponding to modes with low contribution to determine whether there is data acquisition deviation or noise interference. S73. An adversarial sample detection mechanism is introduced to generate a small feature perturbation for the current detection result, and then input it again into the detection algorithm model to verify the robustness of the model's decision. S74. Based on the comprehensive confidence score, confirmatory analysis results, and robustness verification results, the initial detection report is calibrated. When the confidence score after calibration is higher than the preset threshold, the final hardware Trojan detection report is generated. If the value falls below the threshold, an alarm is triggered, and it is recommended to use other alternative algorithms from the standard chip multimodal fusion feature template library matrix for cross-validation.

[0014] Preferably, the method further includes a model continuous optimization step: S81. The chip multimodal fusion feature data, the target hardware Trojan detection algorithm type, and the final hardware Trojan detection report generated in this detection task are desensitized to form a new training sample. S82. Add the new training samples to the background training dataset corresponding to the standard chip multimodal fusion feature template library matrix; S83. Regularly use the updated training dataset to incrementally train and fine-tune the hardware Trojan detection algorithm model.

[0015] Preferably, the system includes a signal sensing and fusion module, a detection strategy matching and execution module, and an evaluation and optimization module; The signal sensing and fusion module acquires static and dynamic side-channel data of the chip under test through a multi-modal signal sensing unit, generates standardized feature data through a signal preprocessing and alignment unit, and outputs multi-modal fusion feature data of the chip through a feature fusion unit. The detection strategy matching and execution module receives the chip's multimodal fusion feature data, calls the standard feature template library through the knowledge base management unit, matches the target detection algorithm type through the strategy decision unit, and outputs a primary hardware Trojan detection report through the detection execution unit. The evaluation and optimization module receives the primary hardware Trojan detection report, calibrates the detection results through a credibility evaluation unit and a robustness verification unit, generates a final hardware Trojan detection report using a feedback optimization unit, and drives continuous updates of system parameters and template library through a model iteration unit.

[0016] Compared with existing technologies, this invention provides a hardware Trojan detection method and system based on multimodal side-channel fusion, which has the following beneficial effects: 1. In this invention, when detecting hardware Trojans, static and dynamic multimodal side-channel data of the chip under test are collected, preprocessed, and feature fused to obtain multimodal fused feature data of the chip. This data is then matched with a pre-built standard feature template library to select target detection algorithms, ultimately completing Trojan identification and location. This invention achieves comprehensive utilization of multimodal side-channel features and intelligent algorithm matching, more comprehensively characterizing chip behavior and improving the detection rate and accuracy of hardware Trojans.

[0017] 2. In this invention, multi-dimensional side-channel data containing static and dynamic characteristics is collected, and data cleaning, frequency alignment, and multi-scale feature extraction are performed. An attention mechanism is used for adaptive weight allocation and fusion to generate chip multi-modal fusion feature data. This achieves in-depth processing and key feature enhancement of multi-source heterogeneous data, ensuring the comprehensiveness and optimization of feature representation, and laying a solid foundation for subsequent high-precision Trojan detection.

[0018] 3. In this invention, by constructing a template library containing standard feature vectors of various detection algorithms and adaptively matching the best detection algorithm, the results are calibrated and cross-validated in multiple dimensions through confidence assessment, contribution analysis, data re-acquisition verification, and adversarial sample robustness testing. This improves the adaptive capability, decision reliability, and resistance to abnormal interference of the detection system. Attached Figure Description

[0019] Figure 1 This is a flowchart of the hardware Trojan detection method based on multimodal side-channel fusion of the present invention; Figure 2 This is a diagram illustrating the architecture of the hardware Trojan detection system based on multimodal side-channel fusion according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For specific implementation examples, please refer to: Figure 1-2 A hardware Trojan detection method and system based on multimodal side-channel fusion, the method including the following steps: S1. Collect static multimodal side-channel data and dynamic multimodal side-channel data of the integrated circuit chip under test; S2. Perform preprocessing and data acquisition frequency alignment processing on the static multimodal side channel data and dynamic multimodal side channel data to generate standardized static multimodal side channel feature data and standardized dynamic multimodal side channel feature data. S3. Perform multi-mode side-channel feature fusion processing based on standardized static multi-mode side-channel feature data and standardized dynamic multi-mode side-channel feature data to generate chip multi-mode fusion feature data; S4. Construct a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms; S5. Match the chip multimodal fusion feature data with the standard chip multimodal fusion feature template library to filter out the target hardware Trojan detection algorithm type; S6. Based on the target hardware Trojan detection algorithm type and chip multimodal fusion feature data, perform hardware Trojan existence determination and location analysis on the integrated circuit chip under test, and generate a primary hardware Trojan detection report. S7. Perform credibility calibration and optimization on the initial hardware Trojan detection report to generate the final hardware Trojan detection report.

[0022] S1 acquires static and dynamic multimode side-channel data of the integrated circuit chip under test, including the following steps: S11. Collect multi-mode side-channel data of the integrated circuit chip under test at the unpowered static operating point using an integrated circuit static parameter analyzer, and generate static multi-mode side-channel data, which includes the chip's static current data, static electromagnetic radiation data, static thermal distribution data and chip die microscopic image data. S12. Collect multi-modal side-channel data of the integrated circuit chip under test under dynamic operation with preset test vectors loaded through the integrated circuit dynamic test platform, and generate dynamic multi-modal side-channel data, which includes dynamic power consumption trajectory data, transient electromagnetic radiation data, runtime sequence data and dynamic thermal imaging data of chip surface.

[0023] The preprocessing and data acquisition frequency alignment process in S2 includes the following steps: S21. Perform data cleaning on the static multimodal side channel data and the dynamic multimodal side channel data respectively, remove outliers and noise, and perform normalization processing to generate preprocessed static multimodal side channel data and preprocessed dynamic multimodal side channel data. In practice, the first step is to clean the collected raw static and dynamic multimodal side-channel data. For each mode's data sequence, outliers are removed using a statistical method, specifically the 3σ principle, where data points deviating more than three standard deviations from the mean are considered outliers and removed or corrected. Subsequently, a digital filter, specifically a low-pass Butterworth filter, is used to remove high-frequency noise. Its transfer function is: ; in For the filter at frequency Frequency response at that point, The cutoff frequency, Here is the filter order; after cleaning, the data is normalized to scale each modal data to the [0,1] interval. The normalization formula is: ; in For the original data points, For the normalized data points, and These are the minimum and maximum values ​​of the modal data sequence, respectively; S22. Identify the data acquisition frequencies of the preprocessed static multimodal side channel data and the preprocessed dynamic multimodal side channel data. If the frequencies are inconsistent, use an interpolation algorithm to adjust the data acquisition frequency of the preprocessed static multimodal side channel data to be consistent with the data acquisition frequency of the preprocessed dynamic multimodal side channel data, and generate frequency-aligned static multimodal side channel data. Interpolation algorithm formula: ; in These are the static side-channel data values ​​obtained through interpolation. , Representing the target time point respectively The values ​​of two adjacent known static data sampling points, This indicates the target time point where static data values ​​need to be inserted. , Corresponding to data value and The original sampling time point; S23. Perform multi-scale feature extraction on the frequency-aligned static multimodal side-channel data and the preprocessed dynamic multimodal side-channel data respectively, extract key features in the time domain, frequency domain and time-frequency domain, and generate standardized static multimodal side-channel feature data and standardized dynamic multimodal side-channel feature data. Time-frequency domain feature extraction is performed using the following transformation formula: ;in This is the original time-domain side-channel signal. Centered on time Window function, For frequency, The imaginary unit, In time and frequency The time-frequency coefficients at that location.

[0024] The multimodal side-channel feature fusion processing in S3 includes the following steps: S31. Concatenate the standardized static multimodal side-channel feature data with the standardized dynamic multimodal side-channel feature data to construct the original multimodal feature vector. S32. An attention mechanism is used to adaptively assign weights to features of different modalities in the original multimodal feature vector, thereby enhancing modal features that are highly correlated with hardware Trojan behavior and suppressing irrelevant and noisy modal features. The formula for calculating attention weights is as follows: ; in For the calculated first Attention weights for each modality feature, with values ​​between 0 and 1. Let i be the eigenvector of the i-th mode. For a globally learnable query vector, For relevance scoring functions, The total number of modalities participating in the fusion. The index variable for summation; S33. Perform dimensionality reduction and fusion on the multimodal features after weight allocation to generate chip multimodal fusion feature data; Feature fusion and dimensionality reduction formulas: ; in This is the final generated chip multimodal fusion feature vector. Let be the attention weight for the i-th modal feature. This is the characteristic transformation function.

[0025] The S4 implementation involves constructing a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms, including the following steps: S41. Establish a multimodal fusion feature template library matrix for standard chips corresponding to different hardware Trojan detection algorithms. , ,in Indicates the first The standard chip multimodal fusion feature vector corresponding to the various hardware Trojan detection algorithm types This represents the maximum number of hardware Trojan detection algorithm types. S42. Hardware Trojan detection algorithm models include classification algorithms based on support vector machines, anomaly detection algorithms based on deep neural networks, template matching algorithms based on side-channel power consumption analysis, and detection algorithms based on timing violation analysis, as well as multimodal fusion feature vectors of standard chips. Indicates the use of the first The standard feature template obtained by the detection algorithm after multimodal side-channel data acquisition, preprocessing, feature extraction and fusion of known Trojan-free gold chips.

[0026] S5 includes the following steps: S51. Calculate the matching degree between the chip multimodal fusion feature data and each standard feature vector in the standard chip multimodal fusion feature template library matrix; The matching degree is calculated using cosine similarity, and the formula is as follows: ; in The matching degree is [-1, 1]. This is the final generated chip multimodal fusion feature vector. This represents the standard chip multimodal fusion feature vector corresponding to the type of hardware Trojan detection algorithm. S52. The improved Osprey optimization algorithm is used to search for the hardware Trojan detection algorithm type corresponding to the standard feature vector with the highest matching degree with the chip multimodal fusion feature data, and this algorithm type is determined as the target hardware Trojan detection algorithm type. The specific search steps are as follows: S521. Initialize the parameters of the improved Osprey optimization algorithm, including the Osprey population size and the maximum number of iterations, and randomly initialize the position of the Osprey population in the algorithm search space, which corresponds to different candidate detection algorithm types. S522, Exploration Phase: Simulate the behavior of ospreys recognizing and exploring prey. Individual ospreys randomly explore different detection algorithm types in the search space and calculate their matching degree with the chip's multimodal fusion feature data as a fitness value. Update the osprey's position based on the fitness value. The position update formula is as follows: ; in , These are the positions of the first individual osprey before and after the update, respectively. This is the best historical location found in the entire current osprey population. This represents a target location randomly selected by the algorithm during the exploration phase. , Represents two independently generated random weight coefficients in the range [0,1); S523. During the development phase, the behavior of ospreys capturing and consuming prey is simulated. Individual ospreys conduct a detailed search in the vicinity of their current location to find a detection algorithm type with better matching degree and update their position accordingly. S524. Iteratively execute the exploration and development phase until the maximum number of iterations is met, and output the hardware Trojan detection algorithm type corresponding to the Osprey position with the highest fitness value, i.e. the best matching degree, as the target hardware Trojan detection algorithm type.

[0027] The determination and location analysis of hardware Trojans in S6 includes the following steps: S61. Combine the chip multimodal fusion feature data with the target hardware Trojan detection algorithm type to construct a hardware Trojan detection task package; S62. Based on the target hardware Trojan detection algorithm type, call the corresponding detection algorithm model; S63. Input the chip's multimodal fusion feature data into the called detection algorithm model, perform inference calculations, and output the preliminary Boolean determination result of the Trojan's existence and the location information of the circuit module where the Trojan is located. S64. Integrate the results of the Trojan existence determination, location information, the type of algorithm used, and a summary of the feature data to generate a basic hardware Trojan detection report.

[0028] The result confidence calibration and optimization process in S7 includes the following steps: S71. Conduct a reliability analysis on the primary hardware Trojan detection report, including calculating the confidence score of the detection results and analyzing the contribution of different modal features to the detection results. The confidence score formula is as follows: ; in The final confidence score is... To determine the degree of matching with the final selected algorithm template, Let the standard deviation of the attention weights for each modality be denoted as . The mean of the attention weights for each modality. , The preset weighting coefficients and ; S72. Based on contribution analysis, re-acquire and verify the original side channel data corresponding to modes with low contribution to determine whether there is data acquisition deviation or noise interference. S73. An adversarial sample detection mechanism is introduced to generate a small feature perturbation for the current detection result, and then input it again into the detection algorithm model to verify the robustness of the model's decision. To verify the robustness of the detection model's decision boundary, an adversarial sample detection mechanism is introduced. Based on the target algorithm model and input features used in this detection, a fast gradient sign method is employed to quickly generate an adversarial perturbation. The perturbation generation formula is: ;in For the generated adversarial perturbation vector, The disturbance intensity coefficient is... For the model loss function Regarding input features gradient, To detect the algorithm model, This represents the model's prediction of the original input. S74. Based on the comprehensive confidence score, confirmatory analysis results, and robustness verification results, the initial detection report is calibrated. When the confidence score after calibration is higher than the preset threshold, the final hardware Trojan detection report is generated. If the value falls below the threshold, an alarm is triggered, and it is recommended to use other alternative algorithms from the standard chip multimodal fusion feature template library matrix for cross-validation.

[0029] The method also includes a continuous model optimization step: S81. After desensitization processing, the chip multimodal fusion feature data, target hardware Trojan detection algorithm type, and final hardware Trojan detection report generated in this detection task are used to form a new training sample. The de-identification process includes: removing and hashing all metadata fields that could directly or indirectly identify specific chip models, production batch numbers, and affiliated companies; and applying differential privacy technology to the core chip multimodal fusion feature data to add an appropriate amount of random noise that follows a Laplace distribution. The noise addition formula can be expressed as: ; ;in This is the desensitized feature vector after adding noise. This is the original fused feature vector. To obtain the scale parameter is Random noise sampled from the Laplace distribution, Features Sensitivity Budget for privacy; S82. Add the new training samples to the background training dataset corresponding to the standard chip multimodal fusion feature template library matrix; S83. Regularly use the updated training dataset to incrementally train and fine-tune the hardware Trojan detection algorithm model.

[0030] The system includes a signal sensing and fusion module, a detection strategy matching and execution module, and an evaluation and optimization module; The signal sensing and fusion module acquires static and dynamic side-channel data of the chip under test through the multi-modal signal sensing unit, generates standardized feature data through the signal preprocessing and alignment unit, and outputs the chip's multi-modal fused feature data through the feature fusion unit. The detection strategy matching and execution module receives multimodal fusion feature data of the chip, calls the standard feature template library through the knowledge base management unit, matches the target detection algorithm type through the strategy decision unit, and outputs a primary hardware Trojan detection report through the detection execution unit. The evaluation and optimization module receives the initial hardware Trojan detection report, calibrates the detection results through the credibility evaluation unit and the robustness verification unit, generates the final hardware Trojan detection report using the feedback optimization unit, and drives the continuous updating of system parameters and template library through the model iteration unit.

[0031] The operation steps of this method and system are as follows: First, multimodal side-channel data of the integrated circuit chip under test (ICD) in both static (power-off) and dynamic operating states are collected using an integrated circuit static parameter analyzer and a dynamic test platform. This includes static current, electromagnetic radiation, thermal distribution, microscopic images, dynamic power consumption trajectory, transient electromagnetic radiation, operating sequence, and dynamic thermal imaging data. Next, these static and dynamic multimodal side-channel data undergo data cleaning and normalization preprocessing. Interpolation algorithms are used to align the data acquisition frequencies, followed by multi-scale feature extraction to generate standardized static and dynamic multimodal side-channel feature data. Subsequently, the standardized feature data are concatenated to construct the original multimodal feature vector. An attention mechanism is used to adaptively weight different modal features to enhance key features and suppress noise. Finally, the weighted features are dimensionality reduced and fused to generate multimodal fused feature data for the chip.

[0032] The system pre-constructs a standard chip multimodal fusion feature template library containing various hardware Trojan detection algorithms. Then, it calculates the matching degree between the real-time generated chip multimodal fusion feature data and each standard feature vector in the template library, and uses an improved Osprey optimization algorithm to search for the most suitable target hardware Trojan detection algorithm type. Next, the system calls the corresponding detection algorithm model based on the target algorithm type, inputs the chip multimodal fusion feature data into the model for inference calculation, outputs a Boolean judgment result on the existence of the Trojan and its location information in the circuit module, and integrates these to generate a preliminary hardware Trojan detection report.

[0033] Finally, the initial report undergoes result credibility calibration and optimization. This includes calculating the confidence score of the detection results, analyzing the contribution of different modal features, performing confirmatory re-collection of the original data corresponding to modalities with low contribution, and introducing an adversarial sample detection mechanism to verify the robustness of the model's decision-making. The results are then calibrated by combining all factors. If the confidence score is higher than a threshold, a final hardware Trojan detection report is generated; if it is lower than the threshold, an alarm is triggered, and it is recommended to use other algorithms from the template library for cross-validation. Furthermore, the system includes a continuous model optimization step, where the data and conclusions from each detection are anonymized and added to the background training dataset for periodic incremental training and fine-tuning of the detection algorithm model.

[0034] The entire system is implemented through the collaborative efforts of the signal sensing and fusion module, the detection strategy matching and execution module, and the evaluation and optimization module.

[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hardware Trojan detection method based on multimodal side-channel fusion, characterized in that: The method includes the following steps: S1. Collect static multimodal side-channel data and dynamic multimodal side-channel data of the integrated circuit chip under test; S2. Preprocess and align the static multimodal side channel data and dynamic multimodal side channel data with the data acquisition frequency to generate standardized static multimodal side channel feature data and standardized dynamic multimodal side channel feature data. S3. Perform multi-mode side-channel feature fusion processing based on the standardized static multi-mode side-channel feature data and the standardized dynamic multi-mode side-channel feature data to generate chip multi-mode fusion feature data; S4. Construct a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms; S5. Match the chip multimodal fusion feature data with the standard chip multimodal fusion feature template library to filter out the target hardware Trojan detection algorithm type; S6. Based on the target hardware Trojan detection algorithm type and the chip multimodal fusion feature data, perform hardware Trojan existence determination and location analysis on the integrated circuit chip under test, and generate a primary hardware Trojan detection report. S7. Perform result credibility calibration and optimization on the primary hardware Trojan detection report to generate the final hardware Trojan detection report.

2. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 1, characterized in that: The step S1, which involves acquiring static and dynamic multimode side-channel data of the integrated circuit chip under test, includes the following steps: S11. Collect multi-mode side-channel data of the integrated circuit chip under test at the unpowered static operating point using an integrated circuit static parameter analyzer, and generate static multi-mode side-channel data, wherein the static multi-mode side-channel data includes the chip's static current data, static electromagnetic radiation data, static thermal distribution data, and chip die microscopic image data. S12. Collect multi-modal side-channel data of the integrated circuit chip under test under dynamic operation with preset test vectors loaded through the integrated circuit dynamic test platform, and generate dynamic multi-modal side-channel data. The dynamic multi-modal side-channel data includes dynamic power consumption trajectory data, transient electromagnetic radiation data, runtime sequence data and dynamic thermal imaging data of the chip surface.

3. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 2, characterized in that: The preprocessing and data acquisition frequency alignment process in S2 includes the following steps: S21. Perform data cleaning on the static multimodal side channel data and the dynamic multimodal side channel data respectively, remove outliers and noise, and perform normalization processing to generate preprocessed static multimodal side channel data and preprocessed dynamic multimodal side channel data. S22. Identify the data acquisition frequencies of the preprocessed static multimodal side channel data and the preprocessed dynamic multimodal side channel data. If the frequencies are inconsistent, use an interpolation algorithm to adjust the data acquisition frequency of the preprocessed static multimodal side channel data to be consistent with the data acquisition frequency of the preprocessed dynamic multimodal side channel data, and generate frequency-aligned static multimodal side channel data. S23. Perform multi-scale feature extraction on the frequency-aligned static multimodal side-channel data and the preprocessed dynamic multimodal side-channel data respectively, extract key features in the time domain, frequency domain, and time-frequency domain, and generate standardized static multimodal side-channel feature data and standardized dynamic multimodal side-channel feature data.

4. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 3, characterized in that: The multimodal side-channel feature fusion processing in S3 includes the following steps: S31. Concatenate the standardized static multimodal side-channel feature data with the standardized dynamic multimodal side-channel feature data to construct the original multimodal feature vector. S32. An attention mechanism is used to adaptively assign weights to the features of different modalities in the original multimodal feature vector, thereby enhancing the modal features that are highly correlated with the behavior of the hardware Trojan and suppressing irrelevant and noisy modal features. S33. Perform dimensionality reduction and fusion on the multimodal features after weight allocation to generate chip multimodal fusion feature data.

5. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 4, characterized in that: The step S4 involves constructing a standard chip multimodal fusion feature template library corresponding to different hardware Trojan detection algorithms, including the following steps: S41. Establish a multimodal fusion feature template library matrix for standard chips corresponding to different hardware Trojan detection algorithms. , ,in Indicates the first The standard chip multimodal fusion feature vector corresponding to the various hardware Trojan detection algorithm types This represents the maximum number of hardware Trojan detection algorithm types. S42. The hardware Trojan detection algorithm model types include classification algorithms based on support vector machines, anomaly detection algorithms based on deep neural networks, template matching algorithms based on side-channel power consumption analysis, and detection algorithms based on timing violation analysis. The standard chip multimodal fusion feature vector... Indicates the use of the first The standard feature template obtained by the detection algorithm after multimodal side-channel data acquisition, preprocessing, feature extraction and fusion of known Trojan-free gold chips.

6. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 5, characterized in that: S5 includes the following steps: S51. Calculate the matching degree between the chip multimodal fusion feature data and each standard feature vector in the standard chip multimodal fusion feature template library matrix; S52. The improved Osprey optimization algorithm is used to search for the hardware Trojan detection algorithm type corresponding to the standard feature vector with the highest matching degree with the chip multimodal fusion feature data, and this algorithm type is determined as the target hardware Trojan detection algorithm type. The specific search steps are as follows: S521. Initialize the parameters of the improved Osprey optimization algorithm, including the Osprey population size and the maximum number of iterations, and randomly initialize the position of the Osprey population in the algorithm search space, which corresponds to different candidate detection algorithm types. S522, Exploration Phase: Simulate the behavior of ospreys recognizing and exploring prey. Individual ospreys randomly explore different detection algorithm types in the search space and calculate their matching degree with the chip's multimodal fusion feature data as a fitness value. Update the osprey's position based on the fitness value. S523. During the development phase, the behavior of ospreys capturing and consuming prey is simulated. Individual ospreys conduct a detailed search in the vicinity of their current location to find a detection algorithm type with better matching degree and update their position accordingly. S524. Iteratively execute the exploration and development phase until the maximum number of iterations is met, and output the hardware Trojan detection algorithm type corresponding to the Osprey position with the highest fitness value, i.e. the best matching degree, as the target hardware Trojan detection algorithm type.

7. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 6, characterized in that: The hardware Trojan existence determination and location analysis in S6 includes the following steps: S61. Combine the chip multimodal fusion feature data with the target hardware Trojan detection algorithm type to construct a hardware Trojan detection task package; S62. Based on the target hardware Trojan detection algorithm type, call the corresponding detection algorithm model; S63. Input the chip's multimodal fusion feature data into the called detection algorithm model, perform inference calculations, and output the preliminary Boolean determination result of the Trojan's existence and the location information of the circuit module where the Trojan is located. S64. Integrate the results of the Trojan existence determination, location information, the type of algorithm used, and a summary of the feature data to generate a basic hardware Trojan detection report.

8. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 7, characterized in that: The result reliability calibration and optimization process in S7 includes the following steps: S71. Perform a reliability analysis on the primary hardware Trojan detection report, including calculating the confidence score of the detection results and analyzing the contribution of different modal features to the detection results. S72. Based on contribution analysis, re-acquire and verify the original side channel data corresponding to modes with low contribution to determine whether there is data acquisition deviation or noise interference. S73. An adversarial sample detection mechanism is introduced to generate a small feature perturbation for the current detection result, and then input it again into the detection algorithm model to verify the robustness of the model's decision. S74. Based on the comprehensive confidence score, confirmatory analysis results, and robustness verification results, the initial detection report is calibrated. When the confidence score after calibration is higher than the preset threshold, the final hardware Trojan detection report is generated. If the value falls below the threshold, an alarm is triggered, and it is recommended to use other alternative algorithms from the standard chip multimodal fusion feature template library matrix for cross-validation.

9. The hardware Trojan detection method based on multimodal side-channel fusion according to claim 8, characterized in that: The method also includes a continuous model optimization step: S81. The chip multimodal fusion feature data, the target hardware Trojan detection algorithm type, and the final hardware Trojan detection report generated in this detection task are desensitized to form a new training sample. S82. Add the new training samples to the background training dataset corresponding to the standard chip multimodal fusion feature template library matrix; S83. Regularly use the updated training dataset to incrementally train and fine-tune the hardware Trojan detection algorithm model.

10. A hardware Trojan detection system based on multimodal side-channel fusion, used to implement the hardware Trojan detection method based on multimodal side-channel fusion as described in any one of claims 1-9, characterized in that: The system includes a signal sensing and fusion module, a detection strategy matching and execution module, and an evaluation and optimization module; The signal sensing and fusion module acquires static and dynamic side-channel data of the chip under test through a multi-modal signal sensing unit, generates standardized feature data through a signal preprocessing and alignment unit, and outputs multi-modal fusion feature data of the chip through a feature fusion unit. The detection strategy matching and execution module receives the chip's multimodal fusion feature data, calls the standard feature template library through the knowledge base management unit, matches the target detection algorithm type through the strategy decision unit, and outputs a primary hardware Trojan detection report through the detection execution unit. The evaluation and optimization module receives the primary hardware Trojan detection report, calibrates the detection results through a credibility evaluation unit and a robustness verification unit, generates a final hardware Trojan detection report using a feedback optimization unit, and drives continuous updates of system parameters and template library through a model iteration unit.

Images