Method, device and equipment for detecting integrated circuit hardware Trojan based on bypass information fusion

By combining power consumption and electromagnetic information feature fusion detection methods, the problem of difficult detection of hidden hardware Trojans in integrated circuits has been solved, achieving efficient and accurate hardware Trojan identification and ensuring information system security.

CN123069366APending Publication Date: 2026-11-10CHINA ELECTRONICS RELIABILITY AND ENVIRONMENTAL TESTING INSTITUTE ((THE FIFTH INSTITUTE OF ELECTRONICS MINISTRY OF INDUSTRY AND INFORMATION TECHNOLOGY) (CHINA SAIBAO LABORATORY)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611137760.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-29
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively detect hidden hardware trojans in integrated circuits. Conventional functional tests are prone to missing detections, and hardware trojans are difficult to activate under specific conditions.

Method used

By combining power consumption and electromagnetic bypass information, a pre-trained feature extraction model is used to perform multi-dimensional feature extraction and length adjustment. After feature fusion, a classification model is used for detection to identify hardware Trojans in integrated circuits.

Benefits of technology

It improves the coverage and accuracy of hardware Trojan detection, reduces the false negative rate, and can identify hidden integrated circuit Trojans, thus ensuring the security of information systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069366A_ABST
    Figure CN123069366A_ABST
Patent Text Reader

Abstract

The application relates to a bypass information fusion-based integrated circuit hardware Trojan detection method, device and equipment. The method comprises the following steps: obtaining power consumption information and electromagnetic information of a target chip; performing multi-dimensional feature extraction processing on the power consumption information by using a pre-trained feature extraction model to obtain intermediate power consumption features, adjusting the feature length of the intermediate power consumption features to obtain target power consumption features, and performing multi-dimensional feature extraction processing on the electromagnetic information to obtain intermediate electromagnetic features, adjusting the feature length of the intermediate electromagnetic features to obtain target electromagnetic features; determining fusion features according to the target power consumption features and the target electromagnetic features; performing detection processing on the fusion features by using a pre-trained classification model to obtain a detection result; wherein the detection result is used to represent whether an integrated circuit hardware Trojan exists in the target chip. The method can cover multiple types of integrated circuit hardware Trojans, has high detection efficiency, and provides a front-end guarantee for information system security.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of integrated circuit design and manufacturing security technology, and in particular to a method, apparatus and equipment for detecting integrated circuit hardware Trojans based on bypass information fusion. Background Technology

[0002] With the separation of integrated circuit design and manufacturing and the use of third-party IP cores, every stage of chip manufacturing faces the risk of malicious tampering. Hardware Trojans, malicious circuits that can be implanted into the original circuitry by attackers, can not only alter the chip's function and performance but also leak information and render it inoperable. When integrated circuits implanted with hardware Trojans are used in high-security fields such as military systems and finance, they pose a significant security threat. Therefore, testing chips before use and identifying and removing those implanted with hardware Trojans is crucial for ensuring security in these fields.

[0003] In traditional technologies, functional testing can be used to test integrated circuits. This method first automatically generates test vectors, applies them to the chip's inputs to obtain test outputs, and then determines whether the test outputs conform to specifications to detect hardware malware.

[0004] However, because hardware Trojans are usually highly stealthy and can only change their output state under specific non-functional triggering conditions, conventional functional tests are unlikely to activate such hardware Trojans, resulting in a high risk of missed detection. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, and device for detecting integrated circuit hardware Trojans based on bypass information fusion to address the aforementioned technical problems. By simultaneously processing both power consumption and electromagnetic bypass information, it can cover multiple types of hardware Trojans, such as logic gate addition and / or reduction, and layout modification. It features broad coverage of hardware Trojan types and high detection efficiency, enabling the removal of hardware Trojan chips before chip application and ensuring the security of information systems.

[0006] Firstly, this application provides a method for detecting integrated circuit hardware Trojans based on bypass information fusion, including:

[0007] Obtain the power consumption and electromagnetic information of the target chip;

[0008] The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature length of the intermediate power consumption features is then adjusted to obtain the target power consumption features. Similarly, the electromagnetic information is processed by multi-dimensional feature extraction to obtain intermediate electromagnetic features. The feature length of the intermediate electromagnetic features is then adjusted to obtain the target electromagnetic features.

[0009] The fusion characteristics are determined based on the target power consumption characteristics and the target electromagnetic characteristics;

[0010] A pre-trained classification model is used to detect the fused features to obtain detection results; these results are used to characterize whether an integrated circuit Trojan exists in the target chip.

[0011] In one embodiment, the feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer. The feature length of intermediate power consumption features is adjusted to obtain the target power consumption features, including:

[0012] The intermediate power consumption feature is input into the pyramid pooling layer to extract local waveform information at different time scales, and the local waveform information at multiple time scales is spliced ​​together to obtain the first power consumption feature.

[0013] The first power consumption feature is input into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature;

[0014] The second power consumption feature is input into the feature distribution adjustment layer for distribution normalization processing to obtain the target power consumption feature after feature length adjustment.

[0015] In one embodiment, the dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer. A first power consumption feature is input to the dimensionality reduction layer for dimensionality reduction processing to obtain a second power consumption feature, including:

[0016] The first power consumption feature is input into the first fully connected layer for feature dimension compression and dimensionality reduction to obtain the power consumption feature after one dimensionality reduction.

[0017] The power consumption feature after one dimensionality reduction is input into the dropout layer for random deactivation to obtain the power consumption feature after random deactivation.

[0018] The power consumption features after random deactivation are input into the second fully connected layer for feature dimension compression to obtain the second power consumption features.

[0019] In one embodiment, the feature distribution adjustment layer includes a batch normalization and regularization layer. The second power consumption feature is input to the feature distribution adjustment layer to obtain the target power consumption feature after feature length adjustment, including:

[0020] The second power consumption feature is input into the batch normalization layer for normalization processing to obtain the power consumption feature after distribution normalization.

[0021] The power consumption features after distribution normalization are input into the regularization layer for regularization constraints to obtain the target power consumption features after feature length adjustment.

[0022] In one embodiment, a pre-trained feature extraction model is used to perform multi-dimensional feature extraction processing on the power consumption information to obtain intermediate power consumption features, including:

[0023] Set the model parameters of the feature extraction model based on power consumption information;

[0024] The power consumption information is input into the feature extraction model with pre-set model parameters and subjected to multiple convolution operations to obtain intermediate power consumption features.

[0025] In one embodiment, the training process of the feature extraction model includes:

[0026] Obtain the sample power consumption set and the sample electromagnetic set;

[0027] The basic extraction model is trained based on the sample power consumption to obtain the feature extraction model, where the model parameters in the feature extraction model are used to extract power consumption features.

[0028] The basic extraction model is trained based on the sample electromagnetic set to obtain the feature extraction model, where the model parameters in the feature extraction model are used to extract electromagnetic features.

[0029] In one embodiment, determining the fusion characteristics based on the target power consumption characteristics and the target electromagnetic characteristics includes:

[0030] Based on the pre-established power consumption-electromagnetic correspondence, the target power consumption feature is mapped into a query vector, and the target electromagnetic feature is mapped into a key vector and a value vector.

[0031] Based on the similarity between the query vector and the key vector, the attention weight between the target power consumption feature and the target electromagnetic feature is determined;

[0032] The value vectors are weighted and fused according to the attention weights to obtain the fused features.

[0033] In one embodiment, acquiring the power consumption information and electromagnetic information of the target chip includes:

[0034] The voltage signal of the sampling resistor in the target chip is obtained and used as power consumption information; the sampling resistor is connected in series in the power supply circuit of the target chip.

[0035] The electromagnetic field induced voltage signal collected by the magnetic field probe is acquired and used as electromagnetic information; the magnetic field probe is placed on the surface of the target chip.

[0036] Secondly, this application also provides a detection device for integrated circuit hardware Trojans based on bypass information fusion, comprising:

[0037] The information acquisition module is used to acquire the power consumption and electromagnetic information of the target chip.

[0038] The feature extraction module is used to perform multi-dimensional feature extraction processing on power consumption information using a pre-trained feature extraction model to obtain intermediate power consumption features, and to adjust the feature length of the intermediate power consumption features to obtain target power consumption features. It also performs multi-dimensional feature extraction processing on electromagnetic information to obtain intermediate electromagnetic features, and to adjust the feature length of the intermediate electromagnetic features to obtain target electromagnetic features.

[0039] The feature fusion module is used to determine the fused features based on the target power consumption features and the target electromagnetic features;

[0040] The Trojan detection module is used to detect fused features using a pre-trained classification model to obtain detection results; the detection results are used to characterize whether there is an integrated circuit Trojan in the target chip.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0042] Obtain the power consumption and electromagnetic information of the target chip;

[0043] The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature length of the intermediate power consumption features is then adjusted to obtain the target power consumption features. Similarly, the electromagnetic information is processed by multi-dimensional feature extraction to obtain intermediate electromagnetic features. The feature length of the intermediate electromagnetic features is then adjusted to obtain the target electromagnetic features.

[0044] The fusion characteristics are determined based on the target power consumption characteristics and the target electromagnetic characteristics;

[0045] A pre-trained classification model is used to detect the fused features to obtain detection results; these results are used to characterize whether an integrated circuit Trojan exists in the target chip.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0047] Obtain the power consumption and electromagnetic information of the target chip;

[0048] The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature length of the intermediate power consumption features is then adjusted to obtain the target power consumption features. Similarly, the electromagnetic information is processed by multi-dimensional feature extraction to obtain intermediate electromagnetic features. The feature length of the intermediate electromagnetic features is then adjusted to obtain the target electromagnetic features.

[0049] The fusion characteristics are determined based on the target power consumption characteristics and the target electromagnetic characteristics;

[0050] A pre-trained classification model is used to detect the fused features to obtain detection results; these results are used to characterize whether an integrated circuit Trojan exists in the target chip.

[0051] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0052] Obtain the power consumption and electromagnetic information of the target chip;

[0053] The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature length of the intermediate power consumption features is then adjusted to obtain the target power consumption features. Similarly, the electromagnetic information is processed by multi-dimensional feature extraction to obtain intermediate electromagnetic features. The feature length of the intermediate electromagnetic features is then adjusted to obtain the target electromagnetic features.

[0054] The fusion characteristics are determined based on the target power consumption characteristics and the target electromagnetic characteristics;

[0055] A pre-trained classification model is used to detect the fused features to obtain detection results; these results are used to characterize whether an integrated circuit Trojan exists in the target chip.

[0056] The aforementioned method, apparatus, and equipment for detecting integrated circuit hardware Trojans based on bypass information fusion. By separately acquiring the power consumption and electromagnetic information of the target chip, and using a feature extraction model to perform multi-dimensional feature extraction and feature length adjustment on both, this feature extraction scheme can capture the subtle physical layer features generated when the integrated circuit hardware Trojan is in a silent state, effectively avoiding the difficulty in activating hidden Trojans and providing a standardized data foundation for subsequent feature fusion. Furthermore, by determining the fusion features based on the target power consumption and electromagnetic features, the deep nonlinear correlation between power consumption and electromagnetic information can be adaptively explored, realizing the complementary advantages of heterogeneous information. This makes the fusion features more discriminative and robust, solving the problems of single bypass signals being susceptible to environmental noise interference and having insufficient characterization capabilities. Finally, a pre-trained classification model is used to detect the fusion features. This pre-trained classification model detects the fusion features based on the discrimination benchmark learned during the training phase. By judging whether the statistical distribution is abnormal through the discrimination benchmark, any abnormal features that deviate from the norm can be identified, thereby discovering hidden integrated circuit trojans. The above scheme effectively overcomes the problems of strong concealment of integrated circuit hardware trojans, difficulty in activation by conventional functional tests, and high false negative rate. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is an application environment diagram of an integrated circuit hardware Trojan detection method based on bypass information fusion in one embodiment;

[0059] Figure 2 This is a flowchart illustrating a method for detecting integrated circuit hardware Trojans based on bypass information fusion in one embodiment.

[0060] Figure 3 Here is a network structure diagram of a convolutional neural network model in one embodiment;

[0061] Figure 4 This is a schematic diagram of the process of a feature extraction model extracting target power consumption features in one embodiment;

[0062] Figure 5 This is a schematic diagram of the process for dimensionality reduction of the first power consumption feature in one embodiment;

[0063] Figure 6 This is a flowchart illustrating the process of adjusting the feature distribution of the second power consumption feature in one embodiment;

[0064] Figure 7 This is a schematic diagram of a process for multi-dimensional feature extraction of power consumption information using a pre-trained feature extraction model in one embodiment.

[0065] Figure 8 This is a flowchart illustrating the training process of a feature extraction model in one embodiment;

[0066] Figure 9 This is a flowchart illustrating the process of determining fusion features based on target power consumption characteristics and target electromagnetic characteristics in one embodiment.

[0067] Figure 10 This is a schematic diagram of the process for obtaining power consumption and electromagnetic information of a target chip in one embodiment;

[0068] Figure 11 This is a flowchart illustrating a method for detecting integrated circuit hardware Trojans based on bypass information fusion in another embodiment.

[0069] Figure 12 This is a structural block diagram of an integrated circuit hardware Trojan detection device based on bypass information fusion in one embodiment;

[0070] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0072] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0073] The method for detecting integrated circuit hardware Trojans based on bypass information fusion provided in this application embodiment can be applied to, for example... Figure 1 The application environment is shown. Terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed in the cloud or on other network servers. A chip testing site is deployed on terminal 102 to acquire power consumption and electromagnetic information of the target chip when it is operating. A data processing center is deployed on server 104 to receive the power consumption and electromagnetic information, and uses a pre-trained feature extraction model to perform multi-dimensional feature extraction and feature length adjustment on the power consumption and electromagnetic information to obtain target power consumption features and target electromagnetic features; a fusion feature is determined based on the target power consumption and target electromagnetic features; and a pre-trained classification model is used to detect and process the fusion feature to obtain a detection result characterizing whether an integrated circuit hardware Trojan exists in the target chip. Terminal 102 can be, but is not limited to, various personal computers, laptops, tablets, and IoT devices; IoT devices can be smart vehicle devices, etc. Server 104 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0074] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting integrated circuit hardware Trojans based on bypass information fusion is provided, and this method is applied to... Figure 1 Taking the server in the example, the following steps are included:

[0075] Step 201: Obtain the power consumption information and electromagnetic information of the target chip.

[0076] Power consumption information is used to characterize the voltage changes in the power supply circuit of the target chip during operation. Since transistor switching, logic flipping, and the triggering of hardware Trojans within the chip all cause instantaneous changes in power consumption, power consumption information reflects the activity state of the internal circuitry and is a commonly used bypass signal for detecting hardware Trojans. Electromagnetic information is used to characterize the changes in the intensity of the electromagnetic field radiated outward by the target chip during operation. Current changes in different circuit units within the chip generate specific electromagnetic field distributions. Especially when a hardware Trojan is active, its electromagnetic radiation characteristics often differ from the original circuit. Therefore, electromagnetic information is a commonly used bypass signal for detecting hardware Trojans. In this embodiment, the bypass signal refers to the physical information unintentionally leaked by the target chip during normal operation or testing, accompanying circuit switching, logic operations, current flow, and other processes.

[0077] The server receives power consumption and electromagnetic information uploaded by the terminal, which has already been collected and preprocessed at the testing site, thus completing the acquisition of the target chip's power consumption and electromagnetic information. In another implementation, the acquisition terminal deployed at the chip testing site converts the physical quantities generated during the target chip's operation into electrical signals, and then performs analog-to-digital conversion to obtain digitized power consumption and electromagnetic information. The acquisition terminal uploads the digitized signals to the server via the network, and the server receives the signals, thus completing the acquisition of the target chip's power consumption and electromagnetic information.

[0078] Step 202: Use a pre-trained feature extraction model to perform multi-dimensional feature extraction processing on power consumption information to obtain intermediate power consumption features, and adjust the feature length of the intermediate power consumption features to obtain target power consumption features. Also, perform multi-dimensional feature extraction processing on electromagnetic information to obtain intermediate electromagnetic features, and adjust the feature length of the intermediate electromagnetic features to obtain target electromagnetic features.

[0079] Among them, the intermediate power consumption feature is used to characterize the voltage change in the power supply circuit of the target chip during operation, and the intermediate electromagnetic feature is used to characterize the change in the electromagnetic field intensity radiated outward by the target chip during operation; the target power consumption feature is used to characterize the power consumption feature of the target chip after feature length adjustment, and the target electromagnetic feature is used to characterize the electromagnetic feature of the target chip after feature length adjustment; feature length adjustment refers to the normalization of the intermediate power consumption feature and the intermediate electromagnetic feature, which aims to eliminate the feature scale differences caused by different signal sources and different acquisition batches.

[0080] The acquired power consumption information is input into a pre-trained feature extraction model. Deep correlation features of the power consumption information are extracted through multi-layer nonlinear transformations to obtain intermediate power consumption features. Feature length processing is performed on the intermediate power consumption features to remove redundant information and scale bias, resulting in the target power consumption features. Similarly, the acquired electromagnetic information is input into a pre-trained feature extraction model. Deep correlation features of the electromagnetic information are extracted through multi-layer nonlinear transformations to obtain intermediate power consumption features. Feature length processing is performed on the intermediate electromagnetic features to remove redundant information and scale bias, resulting in the target electromagnetic features.

[0081] In one embodiment, the feature extraction model is a Convolutional Neural Network (CNN) model. Figure 3 This is a diagram of the network structure of the convolutional neural network model.

[0082] It should be noted that the input data for the feature extraction model are the power consumption information and electromagnetic information of the target chip, respectively. Both models use the same convolutional neural network structure, but the model parameters of the convolutional neural network are trained separately based on the input power consumption and electromagnetic information.

[0083] For example, power consumption information is input into a pre-trained feature extraction model; after multi-layer nonlinear transformation by the pre-trained feature extraction model, intermediate power consumption features are output; the feature length of the intermediate power consumption features is adjusted to map them to target power consumption features with fixed length and normalized numerical distribution. Electromagnetic information is input into the same feature extraction model; after multi-layer nonlinear transformation by the pre-trained feature extraction model, intermediate electromagnetic features are output; the feature length of the intermediate electromagnetic features is adjusted to map them to target electromagnetic features with the same length and similar distribution characteristics as the target power consumption features.

[0084] The process of extracting features and adjusting feature lengths for power consumption and electromagnetic information not only significantly reduces the computational complexity of subsequent calculations, but also preserves core information that is highly relevant to hardware Trojan behavior, laying a standardized data foundation for subsequent fusion.

[0085] Step 203: Determine the fusion characteristics based on the target power consumption characteristics and the target electromagnetic characteristics.

[0086] Among them, the fusion feature is used to characterize the comprehensive features of the target chip after cross-modal interaction of the target power consumption feature and the target electromagnetic feature. This feature has both the voltage change information on the power supply circuit contained in the target power consumption feature and the electromagnetic field strength change information contained in the target electromagnetic feature, which makes up for the perception blind spot of a single bypass signal and can comprehensively reflect the comprehensive operating status of the internal circuit units of the target chip.

[0087] After obtaining the target power consumption characteristics and target electromagnetic characteristics, this embodiment of the application performs joint modeling of the target power consumption characteristics and target electromagnetic characteristics in a high-level feature space. By mining the inherent relationship between the two in the feature dimension, the information complementarity and enhancement are achieved to determine the fused characteristics.

[0088] Step 204: Use a pre-trained classification model to detect the fused features and obtain the detection results.

[0089] The detection results are used to characterize whether there is an integrated circuit hardware Trojan in the target chip.

[0090] The classification model is used to determine the boundaries of the input fused features. This boundary determination refers to the decision-making process of mapping the fused features to normal or abnormal labels based on the geometric determination criteria of the feature space. In the embodiments of this application, the normal label refers to no Trojan horse, and the abnormal label refers to Trojan horse.

[0091] The fused features of the target chip are input into a pre-trained classification model. This model determines the positional relationship of the fused features of the target chip relative to a preset boundary by comparing the degree of offset between the fused features of the target chip and the pre-trained distribution of malware-free features. In this embodiment, the preset boundary refers to the region of malware-free features in the feature space, with normal labels inside the region and abnormal labels outside the region. If the fused features of the target chip are within the preset boundary, a normal label indicating no malware is output; if the fused features of the target chip are outside the preset boundary, an abnormal label indicating malware is output, thereby obtaining the detection result of the target chip.

[0092] In one embodiment, the classification model is a Support Vector Machine (SVM) model.

[0093] The integrated circuit hardware Trojan detection method based on bypass information fusion provided in this application detects integrated circuit hardware Trojans by simultaneously collecting power consumption and electromagnetic information of the target chip, effectively overcoming the shortcomings of single bypass signals being susceptible to environmental noise interference and having low signal-to-noise ratio. Furthermore, by fusing power consumption and electromagnetic bypass signals to detect integrated circuit hardware Trojans on the target chip, the operating state of the chip can be comprehensively characterized from two dimensions: voltage changes and electromagnetic field strength changes. This not only broadens the coverage of integrated circuit hardware Trojan detection but also enhances the sensitivity and detection capability for small-scale Trojans. Based on this, a pre-trained classification model is used to detect the fused features. This model, based on the discrimination benchmark learned through training, identifies anomalies by comparing the degree of deviation between the fused features and normal feature patterns, thereby discovering hidden Trojans and solving the risk of missed detection caused by the difficulty in activating Trojans in conventional functional testing.

[0094] In one exemplary embodiment, such as Figure 4 As shown, the feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer. The intermediate power consumption features are adjusted in length to obtain the target power consumption features. The model includes the following steps:

[0095] Step 301: Input the intermediate power consumption feature into the pyramid pooling layer, extract local waveform information at different time scales, and stitch together the local waveform information at multiple time scales to obtain the first power consumption feature.

[0096] The Spatial Pyramid Pooling (SPP) layer is a multi-scale feature aggregation structure that divides the input features into several segments with different granularities in the time domain and performs pooling operations on the feature information within each segment. This transforms intermediate features with arbitrary time domain lengths into feature vectors with fixed dimensions, addressing the sensitivity of feature extraction models to input signal length. The first power consumption feature represents the high-dimensional power consumption feature after multi-scale feature concatenation. It not only preserves the key waveform structure of power consumption information but also possesses a fixed length characteristic suitable for subsequent feature processing modules. In this embodiment, the SPP layer corresponds to... Figure 3 Mid-space pyramid pooling (1, 2, 4, 8).

[0097] In this embodiment, a spatial pyramid pooling layer is set in the feature extraction model. Because the data lengths of power consumption and electromagnetic information collected from different target chips vary due to differences in sampling rate settings and test vector lengths, the data length input to the feature extraction model is uncertain. To address the sensitivity of the feature extraction model to input size, this embodiment employs a spatial pyramid pooling strategy to perform multi-level partitioning and pooling of the intermediate power consumption features output from the convolutional layer in the time domain. The spatial pyramid pooling layer divides the intermediate power consumption features into segments 1, 2, 4, and 8, resulting in four feature segments at four time scales. For each feature segment, the spatial pyramid pooling layer performs pooling processing on the feature information within that segment, extracting local waveform features at the corresponding time scale. Finally, the pooling results from different time scales are concatenated in the feature dimension to obtain the first power consumption feature.

[0098] Step 302: Input the first power consumption feature into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature.

[0099] Among them, the dimensionality reduction layer refers to a feature processing structure used to compress feature dimensions, reduce data redundancy, and suppress the risk of overfitting; the second power consumption feature represents the low-dimensional power consumption feature after dimensionality reduction.

[0100] The first power consumption feature is input into the dimensionality reduction layer. The first power consumption feature is initially reduced in dimensionality through high-dimensional feature transformation to extract core semantic information. Overfitting is suppressed by randomly masking the output of some neurons to obtain the power consumption feature after random deactivation. The power consumption feature after random deactivation is input into the low-dimensional feature transformation unit in the dimensionality reduction layer to further compress the feature dimension and obtain the second power consumption feature with lower dimension and stronger discriminative power.

[0101] Step 303: Input the second power consumption feature into the feature distribution adjustment layer for distribution normalization processing to obtain the target power consumption feature after feature length adjustment.

[0102] Among them, the feature distribution adjustment layer refers to a post-processing structure used to calibrate and constrain the feature distribution.

[0103] The second power consumption feature is input into the feature distribution adjustment layer, which maps the data of each dimension of the second power consumption feature to a standard distribution interval with zero mean and unit variance, so as to eliminate the numerical imbalance caused by the difference in units between different feature dimensions. After the above feature distribution adjustment layer is processed, the target power consumption feature with unchanged feature length but highly consistent distribution characteristics is obtained.

[0104] In one embodiment, such as Figure 3 As shown, the feature extraction model is a convolutional neural network model. The network structure of the convolutional neural network consists of 5 convolutional layers, 1 pyramid pooling layer, and 2 fully connected layers, including: 3*1 convolutional layer (16), batch normalization layer, 5*1 convolutional layer (32), batch normalization layer, 8*1 average pooling layer, 16*1 convolutional layer (64), batch normalization layer, 32*1 convolutional layer (128), batch normalization layer, 4*1 max pooling layer, 64*1 convolutional layer (256), spatial pyramid pooling (1, 2, 4, 8), fully connected layer (512), dropout, fully connected layer (128), batch normalization layer, L2 regularization, and 128-dimensional features.

[0105] It should be noted that the processing flow for electromagnetic information analysis in this application embodiment is exactly the same as that for processing power consumption information. The intermediate electromagnetic features are input into a pyramid pooling layer to extract local waveform information at different time scales, and the local waveform information at multiple time scales is spliced ​​together to obtain the first electromagnetic feature. The first electromagnetic feature is then input into a dimensionality reduction layer for dimensionality reduction processing to obtain the second electromagnetic feature. The second electromagnetic feature is then input into a feature distribution adjustment layer for distribution normalization processing to obtain the target electromagnetic feature with adjusted feature length.

[0106] In this embodiment, the pyramid pooling layer adapts to inputs of power consumption and electromagnetic information of different lengths, solving the problem of traditional convolutional neural networks being sensitive to the length of input information; the dimensionality reduction layer removes redundant information and suppresses overfitting, enhancing the model's generalization ability; and the feature distribution adjustment layer unifies the feature scale. The above processing flows complement each other, jointly achieving efficient conversion of bypass signals from raw time-domain data to highly robust features, effectively improving the accuracy detection rate of hardware Trojans in micro-integrated circuits.

[0107] In one exemplary embodiment, such as Figure 5 As shown, the dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer. The first power consumption feature is input into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature. The process includes the following steps:

[0108] Step 401: Input the first power consumption feature into the first fully connected layer for high-dimensional feature transformation processing to obtain the high-dimensional mapped power consumption feature.

[0109] The first fully connected layer represents a transformation unit used for high-dimensional nonlinear mapping and feature expansion of the input features; the power consumption feature after high-dimensional mapping represents the transition feature obtained after nonlinear transformation. In this embodiment, the first fully connected layer corresponds to... Figure 3 The fully connected layer (512) is in the middle.

[0110] The first power consumption feature is input into the first fully connected layer. The first power consumption feature is linearly combined using the fully connected operation, and a non-linear activation function is introduced to project the feature to a higher-dimensional feature space to decouple feature associations and mine deep semantic information, thereby obtaining the power consumption feature after high-dimensional mapping.

[0111] Step 402: Input the power consumption features after high-dimensional mapping into the dropout layer for random deactivation processing to obtain the power consumption features after random deactivation.

[0112] Here, the dropout layer refers to a random mask structure that prevents the feature extraction model from overfitting; the power consumption feature after random deactivation represents the transition feature obtained after noise suppression and structural perturbation. In the embodiments of this application, the dropout layer corresponds to Figure 3 Dropout in the middle.

[0113] The features extracted by the first fully connected layer have a high dimensionality. Therefore, a dropout layer is used to randomly deactivate some neurons in the high-dimensional features to prevent overfitting of the feature extraction model, thus obtaining the power consumption features after random deactivation. The processed power consumption features after random deactivation retain the core discriminative information while significantly reducing the model's sensitivity to noise and improving the generalization ability of the power consumption features.

[0114] Step 403: Input the power consumption features after random deactivation into the second fully connected layer for feature dimension compression processing to obtain the second power consumption features.

[0115] The second fully connected layer represents the transformation unit used for information filtering and dimensionality compression of power consumption characteristics after random deactivation; the second power consumption feature represents the low-dimensional power consumption characteristics after dimensionality reduction.

[0116] The power consumption features after random deactivation are input into the second fully connected layer. A linear weighted transformation is used to rearrange and filter the information in the high-dimensional feature space, and a nonlinear activation function is introduced to further refine the discriminative boundaries of the features. The aim is to eliminate redundant components and subtle noise remaining after random deactivation, mapping the power consumption features from a higher-dimensional transition space to a lower-dimensional, higher-information-density low-dimensional power consumption feature space, thus obtaining the second power consumption feature.

[0117] It should be noted that the processing flow for electromagnetic information analysis in this application embodiment is exactly the same as that for processing power consumption information. The first electromagnetic feature is input to the first fully connected layer for high-dimensional feature transformation processing to obtain the high-dimensional mapped electromagnetic feature; the high-dimensional mapped electromagnetic feature is input to the dropout layer for random deactivation processing to obtain the randomly deactivated electromagnetic feature; the randomly deactivated electromagnetic feature is input to the second fully connected layer for feature dimension compression processing to obtain the second electromagnetic feature.

[0118] In this embodiment, a first fully connected layer enhances feature representation capabilities, a dropout layer suppresses overfitting to improve generalization, and finally a second fully connected layer achieves feature dimensionality reduction and information filtering. This effectively solves the computational burden and overfitting risk brought about by high-dimensional features.

[0119] In one exemplary embodiment, such as Figure 6 As shown, the feature distribution adjustment layer includes a batch normalization layer and a regularization layer. The second power consumption feature is input into the feature distribution adjustment layer to obtain the target power consumption feature after feature length adjustment, including the following steps:

[0120] Step 501: Input the second power consumption feature into the batch normalization layer for normalization processing to obtain the power consumption feature after distribution normalization.

[0121] Among them, the batch normalization layer refers to a normalization structure that dynamically calibrates the feature distribution; the power consumption feature representation after distribution normalization is a stable feature representation with zero mean and unit variance after statistical standardization.

[0122] The second power consumption feature is input into the batch normalization layer. First, the data distribution of the second power consumption feature in each dimension is statistically estimated, and the mean and variance of each dimension are calculated. Then, the calculated mean and variance are used to perform linear translation and scaling transformations on the feature in each dimension to make it meet the standard distribution characteristics of zero mean and unit variance. After the above normalization process, the power consumption feature with normalized distribution is obtained.

[0123] Step 502: Input the power consumption features after distribution normalization into the regularization layer for regularization constraint to obtain the target power consumption features after feature length adjustment.

[0124] The regularization layer refers to a post-processing structure used to impose scale normalization constraints on feature vectors.

[0125] In this embodiment of the application, the regularization layer includes L2 regularization, corresponding to Figure 3 L2 regularization is applied. The power consumption feature after distribution normalization is input into the regularization layer, and the L2 norm of the feature is calculated in all dimensions. Then, the L2 norm is used to divide the value of each dimension in the power consumption feature after distribution normalization, so that the L2 norm of the processed feature vector converges to a unit length. After the above processing, the target power consumption feature after feature length adjustment is finally obtained.

[0126] It should be noted that the processing flow for electromagnetic information analysis in this application embodiment is exactly the same as that for processing power consumption information. The second electromagnetic feature is input to the batch normalization layer for normalization processing to obtain the distribution-normalized electromagnetic feature; the distribution-normalized electromagnetic feature is input to the regularization layer for regularization constraint to obtain the target electromagnetic feature with adjusted feature length.

[0127] In this embodiment, by normalizing and regularizing the second power consumption feature, the feature can be standardized in both distribution characteristics and geometric scale, making the distribution of the second power consumption feature extracted from different batches more consistent, and making the final target power consumption feature distribution more stable and consistent in scale.

[0128] In one exemplary embodiment, such as Figure 7 As shown, using a pre-trained feature extraction model to perform multi-dimensional feature extraction on power consumption information to obtain intermediate power consumption features can include the following steps:

[0129] Step 601: Set the model parameters of the feature extraction model based on the power consumption information.

[0130] Before inputting power consumption information into the feature extraction model, model parameters are obtained based on the power consumption information. These model parameters include, but are not limited to, the weights and biases of each neuron.

[0131] It should be noted that the model parameters set by the feature extraction model based on power consumption information may differ from those set based on electromagnetic information.

[0132] Step 602: Input the power consumption information into the feature extraction model with the set model parameters and perform multiple convolution operations to obtain intermediate power consumption features.

[0133] The power consumption information is input into the feature extraction model, where it is sequentially processed through multiple convolutional operations. Utilizing the sliding operation of the convolutional kernels in the time domain, local correlation features and nonlinear patterns are extracted layer by layer from the power consumption information. The final output comprehensively characterizes the intermediate power consumption features that reflect changes in the power supply loop voltage.

[0134] It should be noted that the processing flow for electromagnetic information analysis in this application embodiment is exactly the same as that for processing power consumption information. The model parameters of the feature extraction model are set according to the electromagnetic information; the electromagnetic information is input into the feature extraction model with the set model parameters for multiple convolution operations to obtain intermediate electromagnetic features.

[0135] In this embodiment, by configuring the model parameters of the feature extraction model and using multi-layer convolution operations to deeply mine power consumption information, the process effectively preserves the weak nonlinear features related to the behavior of integrated circuit hardware Trojans in the power consumption information and eliminates redundant noise.

[0136] In one exemplary embodiment, such as Figure 8 As shown, the training process of the feature extraction model includes the following steps:

[0137] Step 701: Obtain the sample power consumption set and the sample electromagnetic set.

[0138] Among them, the sample power consumption set characterizes the voltage change in the power supply circuit of the Trojan-free chip during operation; the sample electromagnetic set characterizes the change in the intensity of the electromagnetic field radiated outward by the Trojan-free chip during operation.

[0139] A Trojan-free chip was selected as the sampling object. The acquisition terminal deployed at the chip testing site monitored the power circuit of the Trojan-free chip and constructed the collected voltage drop or current fluctuation signals into a sample power consumption set. The acquisition terminal deployed at the chip testing site captured the transient magnetic field signal on the surface of the Trojan-free chip package and constructed the collected field strength induced potential into a sample electromagnetic set.

[0140] Step 702: Train the basic extraction model based on the sample power consumption set to obtain the feature extraction model.

[0141] The model parameters in the feature extraction model are used to extract power consumption features. The basic extraction model represents the model trained using the sample power consumption set and the sample electromagnetic set, respectively.

[0142] During the training phase, the sample power consumption set is input into the basic extraction model for training. The model is then trained by sequentially performing feature extraction through convolutional layers, multi-scale aggregation through pyramid pooling layers, feature compression through dimensionality reduction layers, and scale normalization through feature distribution adjustment layers. This results in a well-trained feature extraction model that includes model parameters corresponding to power consumption.

[0143] Step 703: Train the basic extraction model based on the sample electromagnetic set to obtain the feature extraction model.

[0144] Among them, the model parameters in the feature extraction model are used to extract electromagnetic features.

[0145] During the training phase, the sample electromagnetic set is input into the basic extraction model for training. The model is then trained by sequentially performing feature extraction through convolutional layers, multi-scale aggregation through pyramid pooling layers, feature compression through dimensionality reduction layers, and scale normalization through feature distribution adjustment layers. This results in a well-trained feature extraction model that includes model parameters corresponding to the electromagnetic set.

[0146] In this embodiment, power consumption and electromagnetic information of malware-free chips are simultaneously collected to construct sample power consumption sets and sample electromagnetic sets. Subsequently, a basic extraction model with the same structure but independent parameters is trained using the power consumption sample set and the electromagnetic sample set respectively, resulting in a feature extraction model specializing in extracting power consumption and electromagnetic features. This process establishes a feature benchmark for the normal behavior of malware-free chips through unsupervised learning, providing a highly robust feature guarantee for subsequent detection of integrated circuit hardware Trojans based on bypass information fusion.

[0147] In one exemplary embodiment, such as Figure 9 As shown, determining the fusion characteristics based on the target power consumption characteristics and target electromagnetic characteristics may include the following steps:

[0148] Step 801: Based on the pre-established power consumption electromagnetic correspondence, the target power consumption feature is mapped into a query vector, and the target electromagnetic feature is mapped into a key vector and a value vector.

[0149] The relationship between power consumption and queries, as well as the correspondence between electromagnetic properties and keys and values, are established in advance using an attention mechanism. In practical applications, based on these correspondences, the target power consumption characteristics are mapped to queries, and the target electromagnetic characteristics are mapped to keys and values.

[0150] Step 802: Based on the similarity between the query vector and the key vector, determine the attention weight between the target power consumption feature and the target electromagnetic feature.

[0151] By calculating the degree of correlation between the query vector and the key vector in the dot product space or cosine space, the similarity between the target power consumption feature and the electromagnetic features of each dimension is quantified; then the obtained similarity scores are normalized and the values ​​are converted into a probability distribution between 0 and 1 to generate attention weights.

[0152] Step 803: Perform weighted fusion processing on the value vector according to the attention weights to obtain the fused features.

[0153] The attention weights are used as coefficients to perform a weighted summation of the features in each dimension of the value vector, so that the electromagnetic feature components of the target that are highly correlated with the target power consumption feature occupy a larger proportion in the fused features, while the electromagnetic feature components of the target that are less correlated are suppressed. Then, the weighted summation results are superimposed and linearly transformed to obtain the fused features.

[0154] It should be noted that, in the embodiments of this application, the fusion features determined based on the target power consumption features and target electromagnetic features can also be fused by feature addition, feature multiplication, feature splicing, weighted fusion, gated fusion, and bilinear fusion. The embodiments of this application do not limit this, and can be set according to the actual situation.

[0155] In this embodiment, power consumption, electromagnetic interaction and feature reconstruction are realized based on attention weights. This method can adaptively learn the important correlation regions between power consumption and electromagnetic interaction, effectively improving the utilization rate of complementary information.

[0156] In one exemplary embodiment, such as Figure 10 As shown, obtaining the power consumption and electromagnetic information of the target chip can include the following steps:

[0157] Step 901: Obtain the voltage signal of the sampling resistor in the target chip and use the voltage signal as power consumption information.

[0158] The sampling resistor is connected in series in the power supply circuit of the target chip.

[0159] A high-precision sampling resistor is pre-connected in series in the power supply circuit of the target chip, and the voltage signal across the sampling resistor is measured and acquired using an oscilloscope to obtain the power consumption information of the target chip.

[0160] Step 902: Obtain the electromagnetic field induced voltage signal collected by the magnetic field probe, and use the electromagnetic field induced voltage signal as electromagnetic information.

[0161] The magnetic field probe is placed on the surface of the target chip.

[0162] An electric or magnetic field probe is placed on the chip surface beforehand to convert the electromagnetic field into voltage. The voltage output terminal of the probe is connected to an oscilloscope to obtain the electromagnetic information of the chip.

[0163] In this embodiment, a voltage signal representing power consumption information is obtained by connecting a sampling resistor in series in the power supply loop, and an induced voltage signal representing electromagnetic information is obtained by placing a magnetic field probe on the chip surface. The synchronous sampling function of the oscilloscope is used to ensure that the two types of signals are strictly aligned in time, thereby providing an accurate and reliable physical layer data foundation for subsequent fusion analysis based on the feature extraction model.

[0164] In one exemplary embodiment, such as Figure 11 As shown in the figure, this application provides a method for detecting integrated circuit hardware Trojans based on bypass information fusion. The method includes the following steps:

[0165] Step 1001: Obtain the power consumption information and electromagnetic information of the target chip.

[0166] Step 1002: The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer.

[0167] Step 1003: Input the intermediate power consumption feature into the pyramid pooling layer, extract local waveform information at different time scales, and stitch together the local waveform information at multiple time scales to obtain the first power consumption feature.

[0168] Step 1004: Input the first power consumption feature into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature. The dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer.

[0169] Step 1005: Input the second power consumption feature into the feature distribution adjustment layer to obtain the target power consumption feature after feature length adjustment. The feature distribution adjustment layer includes a batch normalization layer and a regularization layer.

[0170] Step 1006: Multidimensional feature extraction is performed on the electromagnetic information using a pre-trained feature extraction model to obtain intermediate electromagnetic features. The feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer.

[0171] Step 1007: Input the intermediate electromagnetic features into the pyramid pooling layer, extract local waveform information at different time scales, and stitch together the local waveform information at multiple time scales to obtain the first electromagnetic features.

[0172] Step 1008: The first electromagnetic feature is input into the dimensionality reduction layer for dimensionality reduction processing to obtain the second electromagnetic feature. The dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer.

[0173] Step 1009: Input the second electromagnetic feature into the feature distribution adjustment layer to obtain the target electromagnetic feature with adjusted feature length. The feature distribution adjustment layer includes batch normalization and regularization layers.

[0174] Step 1010: Determine the fusion characteristics based on the target power consumption characteristics and the target electromagnetic characteristics.

[0175] Step 1011: The fused features are processed using a pre-trained classification model to obtain detection results. These detection results are used to characterize whether an integrated circuit hardware Trojan exists in the target chip.

[0176] In this embodiment, by acquiring two bypass signals—power consumption and electromagnetic signals—and utilizing the non-correlation of these two signals in terms of noise characteristics, environmental interference is significantly suppressed, overcoming the low signal-to-noise ratio of a single signal. A pyramid pooling layer is introduced into the feature extraction model, breaking the dependence of conventional feature extraction models on input length. This allows the feature extraction model to adapt to power consumption and electromagnetic information of different lengths, improving the versatility of the solution. Furthermore, through the feature fusion part, the deep nonlinear correlation between power consumption and electromagnetic information is further explored, achieving complementary advantages of heterogeneous information and making the fused features highly sensitive to small-scale malware. Finally, the fused features are input into a pre-trained classification model for detection processing. The detection results are used to characterize whether an integrated circuit hardware malware exists in the target chip. This pre-trained classification model, based on the discrimination benchmark learned during the training phase, only needs to identify whether there is an abnormal deviation in the distribution of fused features to identify any abnormal risks, fundamentally overcoming the high false negative risk caused by the difficulty in activating hidden malware in conventional functional testing. The solution proposed in this embodiment provides a highly robust and comprehensive technical path for chip security detection.

[0177] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0178] Based on the same inventive concept, this application also provides an integrated circuit hardware Trojan detection device based on bypass information fusion for implementing the aforementioned method for detecting integrated circuit hardware Trojans based on bypass information fusion. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the integrated circuit hardware Trojan detection device based on bypass information fusion provided below can be found in the limitations of the integrated circuit hardware Trojan detection method based on bypass information fusion described above, and will not be repeated here.

[0179] In one exemplary embodiment, such as Figure 12 As shown, a detection device for integrated circuit hardware Trojans based on bypass information fusion is provided, comprising:

[0180] The information acquisition module 1101 is used to acquire the power consumption information and electromagnetic information of the target chip.

[0181] The feature extraction module 1102 is used to perform multi-dimensional feature extraction processing on power consumption information using a pre-trained feature extraction model to obtain intermediate power consumption features, and to adjust the feature length of the intermediate power consumption features to obtain target power consumption features. It also performs multi-dimensional feature extraction processing on electromagnetic information to obtain intermediate electromagnetic features, and to adjust the feature length of the intermediate electromagnetic features to obtain target electromagnetic features.

[0182] The feature fusion module 1103 is used to determine the fused features based on the target power consumption features and the target electromagnetic features.

[0183] Trojan detection module 1104 is used to detect fused features using a pre-trained classification model to obtain detection results; wherein, the detection results are used to characterize whether there is an integrated circuit Trojan in the target chip.

[0184] In one embodiment, the feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer. The feature extraction module 1102 is specifically used to input the intermediate power consumption feature into the pyramid pooling layer, extract local waveform information at different time scales, and splice the local waveform information at multiple time scales to obtain a first power consumption feature; input the first power consumption feature into the dimensionality reduction layer for dimensionality reduction processing to obtain a second power consumption feature; and input the second power consumption feature into the feature distribution adjustment layer for distribution normalization processing to obtain the target power consumption feature after feature length adjustment.

[0185] In one embodiment, the dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer. The feature extraction module 1102 is specifically used to input the first power consumption feature into the first fully connected layer for high-dimensional feature transformation processing to obtain the high-dimensional mapped power consumption feature; input the high-dimensional mapped power consumption feature into the dropout layer for random deactivation processing to obtain the randomly deactivated power consumption feature; and input the randomly deactivated power consumption feature into the second fully connected layer for feature dimension compression processing to obtain the second power consumption feature.

[0186] In one embodiment, the feature distribution adjustment layer includes a batch normalization layer and a regularization layer. The feature extraction module 1102 is specifically used to input the second power consumption feature into the batch normalization layer for normalization processing to obtain the power consumption feature after distribution normalization; and to input the power consumption feature after distribution normalization into the regularization layer for regularization constraint to obtain the target power consumption feature after feature length adjustment.

[0187] In one embodiment, the feature extraction module 1102 is specifically used to set the model parameters of the feature extraction model according to the power consumption information; input the power consumption information into the feature extraction model with the set model parameters to perform multiple convolution operations to obtain intermediate power consumption features.

[0188] In one embodiment, the device further includes:

[0189] The sample acquisition module is used to acquire the sample power consumption set and the sample electromagnetic set;

[0190] The first training module is used to train the basic extraction model based on the sample power consumption set to obtain the feature extraction model. The model parameters in the feature extraction model are used to extract power consumption features.

[0191] The second training module is used to train the basic extraction model based on the sample electromagnetic set to obtain the feature extraction model. The model parameters in the feature extraction model are used to extract electromagnetic features.

[0192] In one embodiment, the feature fusion module 1103 is specifically used to map the target power consumption feature into a query vector and the target electromagnetic feature into a key vector and a value vector according to the pre-established power consumption electromagnetic correspondence; determine the attention weight between the target power consumption feature and the target electromagnetic feature based on the similarity between the query vector and the key vector; and perform weighted fusion processing on the value vector according to the attention weight to obtain the fused feature.

[0193] In one embodiment, the information acquisition module 1101 is specifically used to acquire the voltage signal of the sampling resistor in the target chip and use the voltage signal as power consumption information; the sampling resistor is connected in series in the power supply circuit of the target chip; acquire the electromagnetic field induced voltage signal collected by the magnetic field probe and use the electromagnetic field induced voltage signal as electromagnetic information; the magnetic field probe is disposed on the surface of the target chip.

[0194] The modules in the aforementioned detection device for integrated circuit hardware Trojans based on bypass information fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0195] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores integrated circuit hardware Trojan data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting integrated circuit hardware Trojans based on bypass information fusion.

[0196] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0197] In one exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, executably including a memory of instructions that can be executed by a processor of an electronic device to perform the above-described method. Exemplarily, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0198] In one exemplary embodiment, a computer program product is also provided, which, when executed by a processor, can implement the above-described methods. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, some or all of the above-described methods can be implemented, wholly or partially, according to the processes or functions described in the embodiments of this application.

[0199] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0200] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0201] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0202] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting integrated circuit hardware Trojans based on bypass information fusion, characterized in that, The method includes: Obtain the power consumption and electromagnetic information of the target chip; The power consumption information is processed by multi-dimensional feature extraction using a pre-trained feature extraction model to obtain intermediate power consumption features. The feature length of the intermediate power consumption features is then adjusted to obtain target power consumption features. Similarly, the electromagnetic information is processed by multi-dimensional feature extraction to obtain intermediate electromagnetic features. The feature length of the intermediate electromagnetic features is then adjusted to obtain target electromagnetic features. The fusion characteristics are determined based on the target power consumption characteristics and the target electromagnetic characteristics; The fused features are detected using a pre-trained classification model to obtain detection results; wherein, the detection results are used to characterize whether there is an integrated circuit hardware Trojan in the target chip.

2. The method according to claim 1, characterized in that, The feature extraction model includes a pyramid pooling layer, a dimensionality reduction layer, and a feature distribution adjustment layer. The step of adjusting the feature length of the intermediate power consumption feature to obtain the target power consumption feature includes: The intermediate power consumption feature is input into the pyramid pooling layer to extract local waveform information at different time scales, and the local waveform information at multiple time scales is spliced ​​together to obtain the first power consumption feature. The first power consumption feature is input into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature; The second power consumption feature is input into the feature distribution adjustment layer for distribution normalization processing to obtain the target power consumption feature after feature length adjustment.

3. The method according to claim 2, characterized in that, The dimensionality reduction layer includes a first fully connected layer, a dropout layer, and a second fully connected layer. The step of inputting the first power consumption feature into the dimensionality reduction layer for dimensionality reduction processing to obtain the second power consumption feature includes: The first power consumption feature is input into the first fully connected layer for high-dimensional feature transformation processing to obtain the high-dimensional mapped power consumption feature. The power consumption features after high-dimensional mapping are input into the dropout layer for random deactivation processing to obtain the power consumption features after random deactivation. The power consumption feature after random deactivation is input into the second fully connected layer for feature dimension compression processing to obtain the second power consumption feature.

4. The method according to claim 2, characterized in that, The feature distribution adjustment layer includes a batch normalization layer and a regularization layer. The step of inputting the second power consumption feature into the feature distribution adjustment layer to obtain the target power consumption feature after feature length adjustment includes: The second power consumption feature is input into the batch normalization layer for normalization processing to obtain the power consumption feature after distribution normalization. The power consumption feature after the distribution is normalized is input into the regularization layer for regularization constraint to obtain the target power consumption feature after feature length adjustment.

5. The method according to claim 1, characterized in that, The process of using a pre-trained feature extraction model to perform multi-dimensional feature extraction on the power consumption information to obtain the intermediate power consumption features includes: The model parameters of the feature extraction model are set according to the power consumption information; The power consumption information is input into a feature extraction model with pre-set model parameters and subjected to multiple convolution operations to obtain the intermediate power consumption features.

6. The method according to claim 5, characterized in that, The training process of the feature extraction model includes: Obtain the sample power consumption set and the sample electromagnetic set; The basic extraction model is trained based on the sample power consumption set to obtain the feature extraction model, wherein the model parameters in the feature extraction model are used to extract power consumption features; The basic extraction model is trained based on the sample electromagnetic set to obtain the feature extraction model, wherein the model parameters in the feature extraction model are used to extract electromagnetic features.

7. The method according to any one of claims 1-6, characterized in that, The step of determining the fusion feature based on the target power consumption feature and the target electromagnetic feature includes: Based on the pre-established power consumption electromagnetic correspondence, the target power consumption feature is mapped into a query vector, and the target electromagnetic feature is mapped into a key vector and a value vector. Based on the similarity between the query vector and the key vector, the attention weight between the target power consumption feature and the target electromagnetic feature is determined; The value vector is weighted and fused according to the attention weights to obtain the fused features.

8. The method according to any one of claims 1-6, characterized in that, The acquisition of the power consumption information and electromagnetic information of the target chip includes: The voltage signal of the sampling resistor in the target chip is obtained, and the voltage signal is used as the power consumption information; the sampling resistor is connected in series in the power supply circuit of the target chip; The electromagnetic field induced voltage signal collected by the magnetic field probe is acquired, and the electromagnetic field induced voltage signal is used as the electromagnetic information; the magnetic field probe is set on the surface of the target chip.

9. A detection device for integrated circuit hardware Trojans based on bypass information fusion, characterized in that, The device includes: The information acquisition module is used to acquire the power consumption and electromagnetic information of the target chip. The feature extraction module is used to perform multi-dimensional feature extraction processing on the power consumption information using a pre-trained feature extraction model to obtain intermediate power consumption features, and to adjust the feature length of the intermediate power consumption features to obtain target power consumption features; and to perform multi-dimensional feature extraction processing on the electromagnetic information to obtain intermediate electromagnetic features, and to adjust the feature length of the intermediate electromagnetic features to obtain target electromagnetic features. The feature fusion module is used to determine the fused features based on the target power consumption features and the target electromagnetic features; The Trojan detection module is used to perform detection processing on the fused features using a pre-trained classification model to obtain detection results; wherein, the detection results are used to characterize whether there is an integrated circuit Trojan in the target chip.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.