Hardware Trojan horse detection precision improving method and system and computer readable storage medium
By establishing a process deviation model and combining it with SVR prediction, the detection threshold is dynamically adjusted, which solves the problem of false alarms and false negatives caused by process deviations in hardware Trojan detection, improves the accuracy and reliability of detection, and is applicable to integrated circuits at advanced process nodes.
Patent Information
- Application Number
- CN202511512449.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2026-02-03
AI Technical Summary
Existing hardware Trojan detection methods struggle to accurately distinguish between anomalies caused by process deviations and those caused by hardware Trojans when faced with the impact of process variations, resulting in high false alarm and false negative rates, especially at advanced process nodes where detection becomes more difficult.
By establishing a high-precision process deviation model and combining it with support vector machine regression (SVR) for prediction, the detection threshold is dynamically adjusted to compensate for the impact of process deviation and to distinguish between process deviation and hardware Trojan signals.
It significantly improves the accuracy and reliability of hardware Trojan detection, reduces false alarms and false negatives, and is applicable to large-scale integrated circuits and adapts to changes in different process nodes.
Smart Images
Figure CN121456926A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of integrated circuit security technology, and in particular to a method, system, and computer-readable storage medium for improving the accuracy of hardware Trojan detection. Background Technology
[0002] With the continuous development of integrated circuit manufacturing technology, hardware security has become a major concern. A hardware Trojan is a malicious circuit that may be implanted into an integrated circuit during the design or manufacturing stage, thereby compromising its security, functionality, or reliability. These malicious circuits are typically characterized by high concealment and complex triggering conditions, potentially causing serious consequences such as information leakage, functional tampering, or even system paralysis. Traditional hardware Trojan detection methods mainly rely on side-channel analysis (such as timing and power consumption) and functional testing, but these methods face a significant challenge in practical applications: the impact of process variation.
[0003] Process deviations are variations in chip parameters (such as transistor threshold voltage and channel length) caused by uncertainties in the manufacturing process (such as lithography precision and doping concentration) (Orshansky, 2008). These deviations become increasingly significant as process nodes shrink. Studies have shown that in 45nm and below processes, process deviations can lead to critical path delay variations of up to 15-20% (Bowman, 2009). These deviations can cause the actual chip performance (such as timing and current) to differ from design expectations, thus masking the signals of hardware trojans. For example, process deviations can cause variations in transient supply current and maximum operating frequency, which may resemble anomalies introduced by hardware trojans, severely interfering with detection results.
[0004] In existing technologies, some approaches attempt to statistically analyze the impact of process deviations using Monte Carlo simulation (Singhee & Rutenbar, 2007). While these methods can simulate the statistical characteristics of process deviations, they are typically not integrated with intelligent prediction models, resulting in high computational complexity and difficulty in efficiently applying them to large-scale integrated circuits. Other approaches employ neural networks for error prediction (Liu, 2018), but neural networks suffer from overlearning and underlearning issues, exhibit poor generalization ability in small sample sizes, and have limited adaptability to changes in process corners. Furthermore, these methods often ignore the varying sensitivity of different basic units to process deviations, leading to insufficient compensation accuracy.
[0005] More importantly, most existing process deviation compensation methods operate independently of hardware Trojan detection systems, failing to establish an effective collaborative mechanism. This separate processing approach makes it difficult for the system to accurately distinguish between genuine hardware Trojan signals and noise caused by process deviations in complex process environments. Especially at advanced process nodes, the impact of process deviations and the signal characteristics of miniaturized hardware Trojans often highly overlap, further increasing the difficulty of detection.
[0006] Therefore, there is an urgent need for a method that can accurately quantify process deviations, predict their impact, and effectively compensate for them, in order to enhance the reliability of hardware Trojan detection. An ideal technical solution should possess the following characteristics: (1) the ability to accurately model the impact of process deviations on key circuit parameters; (2) efficient predictive capabilities that can adapt to changes in different process corners; (3) seamless integration with existing hardware Trojan detection systems; and (4) the ability to maintain high prediction accuracy even under small sample conditions (Jouppi, 2020). Such a solution will significantly improve the security protection capabilities of integrated circuits, especially in application scenarios with stringent security requirements. Summary of the Invention
[0007] The purpose of this invention is to provide a modeling and compensation method for process deviations to solve the problems of false positives and false negatives caused by process deviations in existing hardware Trojan detection technologies. By establishing a high-precision process deviation model and combining it with support vector machine regression (SVR) for prediction, this invention can effectively distinguish between anomalies caused by process deviations and anomalies caused by hardware Trojans, thereby improving detection efficiency and accuracy.
[0008] This invention provides a method, system, and computer-readable storage medium for improving the accuracy of hardware Trojan detection. By establishing an accurate process deviation prediction model, it effectively distinguishes between process deviations and hardware Trojan signals, significantly improving detection accuracy. To achieve the above objectives, this invention adopts the following technical solution:
[0009] Methods to improve the accuracy of hardware Trojan detection include the following steps:
[0010] Step 1, Process Deviation Data Acquisition Step: Through Monte Carlo analysis, collect data on multiple key performance parameters of the integrated circuit under different process angles;
[0011] Step 2, Mathematical Modeling Step: Based on the collected data, establish a quantitative relationship model between the process deviation parameters and the key performance parameters through regression analysis;
[0012] Step 3, Machine learning model construction steps: Construct a support vector machine regression model, and use the key performance parameter data collected in Step 1 to train the quantitative relationship model in Step 2 to obtain a prediction model for predicting process deviation parameters.
[0013] Step 4, Compensation Application Step: Integrate the trained prediction model into the hardware Trojan detection system to predict the deviation diffusion factor under the current process conditions in real time, and dynamically adjust the detection threshold to compensate for the impact of process deviation, thereby distinguishing between process deviation anomalies and hardware Trojan anomalies.
[0014] Furthermore, the specific process of step 1 is as follows:
[0015] For the FF, SS, and TT process corners, multiple Monte Carlo random sampling simulations were performed. The simulations included multiple sets of transient supply current Itran and maximum operating frequency Fmax. When determining the standard cell library category and quantity, at least 1000 Monte Carlo random sampling simulations were performed for each set of transient supply current Itran and maximum operating frequency Fmax to obtain time delay data. The time delay data conformed to a Gaussian distribution, and the time delay data at σ=0 in the Gaussian distribution was recorded as the actual cell time delay. The time delay data when σ=50% in the Gaussian distribution is denoted as the threshold time delay. The transient supply current Itran, maximum operating frequency Fmax, and actual unit delay of this group are also considered. and threshold delay This is recorded as a set of key performance parameter data.
[0016] Furthermore, the specific method for step 2 is as follows:
[0017] ;
[0018] The subscripts FF, SS, and TT correspond to three different process angles. , For the corresponding regression coefficients, For process deviation parameters, SC is the One-Hot code corresponding to the standard cell library category and quantity.
[0019] Furthermore, the specific method of step 3 is as follows:
[0020] SVR models were established for the three process corners: FF, SS, and TT. The input for each SVR model consisted of the One-Hot code SC corresponding to the standard cell library category and quantity from Monte Carlo random sampling simulation, along with multiple sets of key performance parameter data corresponding to that SC, used to fit the relationship model in step 2. The transient supply current Itran and the maximum operating frequency Fmax were directly input, and the actual cell delay was used. and threshold delay A sliding window smoothing preprocessing method is adopted; the SVR model uses radial basis function kernels to handle nonlinear relationships; the fault tolerance parameter ε=0.1ns is set to meet the accuracy requirements.
[0021] Furthermore, the SVR model is trained and validated using k-fold cross-validation, where k=5.
[0022] Furthermore, the calculation process for the process parameter deviation is as follows:
[0023] ;
[0024] Keeping the transient supply current Itran, maximum operating frequency Fmax, and standard cell library category constant, the actual cell delay is obtained by changing the number of standard cells. The difference ,as well as The difference .
[0025] Furthermore, step 4 is specifically implemented as follows:
[0026] Threshold_adjusted = Threshold_base × DF;
[0027] Wherein, DF is the process deviation diffusion factor predicted by the SVR model, and its value is: The value corresponding to the process corner; Threshold_base is the original detection threshold in the hardware Trojan detection system, and Threshold_adjusted is the dynamically adjusted detection threshold.
[0028] Furthermore, step 4 also includes security enhancement measures, which include: storing the trained SVR model parameters in a one-time programmable memory; using a hash algorithm to verify the integrity of the model; and setting a hardware watchdog timer to monitor for anomalies in prediction behavior.
[0029] A hardware Trojan detection system for implementing any of the above-described methods for improving the accuracy of hardware Trojan detection includes:
[0030] The data acquisition module is configured to perform the process deviation data acquisition step.
[0031] A modeling and analysis module, configured to perform the mathematical modeling steps;
[0032] The machine learning module is configured to perform the machine learning model building steps and store the trained SVR prediction model.
[0033] The detection and compensation execution module is integrated into the hardware Trojan detection tool. It is configured to load the SVR prediction model and execute the compensation application steps to achieve dynamic adjustment of the detection threshold.
[0034] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements any of the methods described above.
[0035] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art:
[0036] 1. High-precision prediction: It adopts the SVR model to predict the process deviation diffusion factor and solves the problem of nonlinear relationship modeling through kernel function mapping. The prediction accuracy reaches 0.1 nanosecond level (equivalent to 3 FO4 delays in 28nm process). Compared with traditional statistical methods (error is usually >0.5ns), the accuracy is improved by more than 5 times.
[0037] 2. Enhanced detection reliability: The application of compensation parameters enables the detection unit to more accurately distinguish between process deviations and hardware Trojan signals, reducing false alarms and missed alarms.
[0038] 3. Efficiency and adaptability: The combination of Monte Carlo analysis and SVR makes this method applicable to large-scale integrated circuits and to different process nodes (such as nanometer scale and sub-nanometer scale).
[0039] 4. Advantage of small sample size: SVR can still maintain high generalization ability in the case of small sample size, overcome the problem of overfitting in neural networks, and improve the practicality and reliability of the model. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the overall system structure of the present invention.
[0041] Figure 2 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation
[0042] The following description, in conjunction with the accompanying drawings and implementation process, should be understood by those skilled in the art as these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0043] Reference Figure 1 The device of this invention includes: a data acquisition module, a modeling and analysis module, a machine learning module, and a detection and compensation execution module. Wherein:
[0044] The data acquisition module is primarily used for acquiring process deviation data. Monte Carlo analysis tools (such as Cadence Virtuoso) are used to perform transient simulations of typical circuit units (such as inverter chains and ring oscillators), collecting transient supply current and maximum operating frequency data at different process angles (FF, SS, TT, etc.). It is recommended that at least 1000 Monte Carlo simulations be performed to ensure coverage of the statistical distribution of process deviations. Each simulation should record the random variations of process parameters (such as Vth, L) and their impact on electrical parameters.
[0045] The modeling and analysis module is primarily used to execute the mathematical modeling steps. It performs regression analysis on the collected data to establish the functional relationship between process deviations and electrical parameters. For example,
[0046]
[0047] The machine learning module is primarily used to construct the dataset and configure SVM training, execute the machine learning model construction steps, and store the trained SVM prediction model. It extracts features from the Monte Carlo simulation results, including basic cell types (such as NAND, NOR, DFF), cell count, process conditions, and latency diffusion factor. The dataset should contain a sufficient number of samples (at least 10,000 data points are recommended) and be divided into a 70% training set and a 30% test set. Support Vector Machine Regression (SVR) is used for training, and the kernel function can be a radial basis function (RBF). The training objective is to minimize the error between the predicted diffusion factor and the actual diffusion factor.
[0048] SVM Regression (SVR) Model Architecture:
[0049] from sklearn import svm
[0050] model = svm.SVR(kernel='rbf',
[0051] C=1.0,
[0052] epsilon=0.1,
[0053] gamma='scale')
[0054] A radial basis function (RBF) kernel is used to handle nonlinear relationships; a tolerance parameter ε=0.1 is set to meet accuracy requirements. K-fold cross-validation (k=5) is used to prevent overfitting.
[0055] Training process:
[0056] 1) Collect the data required for the above model.
[0057] 2) Establish the SVR model as shown above.
[0058] 3) Training the SVR model
[0059] 3. Diffusion factor extraction
[0060] Diffusion factor matrix:
[0061]
[0062] The diffusion factor k was extracted from the trained SVM model using sensitivity analysis.
[0063]
[0064] Calculated using numerical methods (such as giving SC a small change ΔSC, for example, slightly increasing the number of cells, and observing the change Δy in the SVM output).
[0065]
[0066]
[0067]
[0068] In the above expression Given two inputs that are relatively close to each other, i.e. The difference between the two predictions of the model.
[0069] 4. Application of the process deviation model
[0070] Achieve closed-loop optimization of the testing process.
[0071] Real-time compensation mechanism:
[0072] 1. Loading process feature vectors after chip power-on
[0073] 2. Online prediction of diffusion factor K using the SVR model
[0074] 3. Dynamically adjust the detection threshold:
[0075]
[0076] The detection and compensation execution module, integrated into the hardware Trojan detection tool, is primarily used to configure the loading of the SVM prediction model and the execution of the compensation application, enabling dynamic adjustment of the detection threshold. The trained SVM model can be integrated into the hardware Trojan detection tool. During detection, the model predicts a diffusion factor based on actual process conditions and circuit characteristics. The detection unit uses this diffusion factor to adjust the detection threshold (such as timing deviation threshold and current threshold) to compensate for the impact of process deviations. For example, a hardware Trojan alarm is triggered only when the timing deviation exceeds the prediction range.
[0077] Reference Figure 2 Taking a 65nm chip as an example:
[0078] Data Acquisition: Monte Carlo simulation was performed on the critical path in the chip to acquire 10,000 sets of transient current and frequency data.
[0079] Model training: The model is trained using SVM. Input features include cell type, quantity, and process angle. The output is the time delay diffusion factor.
[0080] Compensation Application: In hardware Trojan detection, the diffusion factor predicted by the model is used to adjust the current threshold. Experiments show that the false alarm rate is reduced by 40%, and the false negative rate is reduced by 25%.
[0081] Through the above implementation methods, the present invention can effectively improve the accuracy and reliability of hardware Trojan detection and is suitable for security testing of integrated circuits at advanced process nodes.
[0082] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various modifications or alterations can be made by those skilled in the art without departing from the spirit and scope of the claims of this application.
Claims
1. A method for improving the accuracy of hardware Trojan detection, characterized in that, Includes the following steps: Step 1, Process Deviation Data Acquisition Step: Through Monte Carlo analysis, collect data on multiple key performance parameters of the integrated circuit under different process angles; Step 2, Mathematical Modeling Step: Based on the collected data, establish a quantitative relationship model between the process deviation parameters and the key performance parameters through regression analysis; Step 3, Machine learning model construction steps: Construct a support vector machine regression model, and use the key performance parameter data collected in Step 1 to train the quantitative relationship model in Step 2 to obtain a prediction model for predicting process deviation parameters. Step 4, Compensation Application Step: Integrate the trained prediction model into the hardware Trojan detection system to predict the deviation diffusion factor under the current process conditions in real time, and dynamically adjust the detection threshold to compensate for the impact of process deviation, thereby distinguishing between process deviation anomalies and hardware Trojan anomalies.
2. The method for improving the accuracy of hardware Trojan detection according to claim 1, characterized in that, The specific process of step 1 is as follows: For the FF, SS, and TT process corners, multiple Monte Carlo random sampling simulations were performed. The simulations included multiple sets of transient supply current Itran and maximum operating frequency Fmax. When determining the standard cell library category and quantity, at least 1000 Monte Carlo random sampling simulations were performed for each set of transient supply current Itran and maximum operating frequency Fmax to obtain delay data. The delay data conformed to a Gaussian distribution, and the delay data at σ=0 in the Gaussian distribution was recorded as the actual cell delay. The time delay data when σ=50% in the Gaussian distribution is denoted as the threshold time delay. The transient supply current Itran, maximum operating frequency Fmax, and actual unit delay of this group are also considered. and threshold delay This is recorded as a set of key performance parameter data.
3. The method for improving the accuracy of hardware Trojan detection according to claim 2, characterized in that, The specific method for step 2 is as follows: ； The subscripts FF, SS, and TT correspond to three different process angles. , For the corresponding regression coefficients, For process deviation parameters, SC is the One-Hot code corresponding to the standard cell library category and quantity.
4. The method for improving the accuracy of hardware Trojan detection according to claim 3, characterized in that, The specific method for step 3 is as follows: SVR models were established for the three process corners: FF, SS, and TT. The input for each SVR model consisted of the One-Hot code SC corresponding to the standard cell library category and quantity from Monte Carlo random sampling simulation, along with multiple sets of key performance parameter data corresponding to that SC, used to fit the relationship model in step 2. The transient supply current Itran and the maximum operating frequency Fmax were directly input, and the actual cell delay was used. and threshold delay A sliding window smoothing preprocessing method is adopted; the SVR model uses radial basis function kernels to handle nonlinear relationships; the fault tolerance parameter ε=0.1ns is set to meet the accuracy requirements.
5. The method for improving the accuracy of hardware Trojan detection according to claim 4, characterized in that, The SVR model was trained and validated using k-fold cross-validation, where k=5.
6. The method according to claim 5, characterized in that, The calculation process for the deviation of the process parameters is as follows: ； Keeping the transient supply current Itran, maximum operating frequency Fmax, and standard cell library category constant, the actual cell delay is obtained by changing the number of standard cells. The difference ,as well as The difference .
7. The method for improving the accuracy of hardware Trojan detection according to claim 1, characterized in that, The specific method for step 4 is as follows: Threshold_adjusted = Threshold_base × DF; Wherein, DF is the process deviation diffusion factor predicted by the SVR model, and its value is: The value corresponding to the process corner; Threshold_base is the original detection threshold in the hardware Trojan detection system, and Threshold_adjusted is the dynamically adjusted detection threshold.
8. The method for improving the accuracy of hardware Trojan detection according to claim 1, characterized in that, Step 4 also includes security enhancement measures, which include: storing the trained SVR model parameters in a one-time programmable memory; using a hash algorithm to verify the integrity of the model; and setting a hardware watchdog timer to monitor for anomalies in prediction behavior.
9. A hardware Trojan detection system for executing the hardware Trojan detection accuracy improvement method as described in any one of claims 1-8, characterized in that, include: The data acquisition module is configured to perform the process deviation data acquisition step. A modeling and analysis module, configured to perform the mathematical modeling steps; The machine learning module is configured to perform the machine learning model building steps and store the trained SVR prediction model. The detection and compensation execution module is integrated into the hardware Trojan detection tool. It is configured to load the SVR prediction model and execute the compensation application steps to achieve dynamic adjustment of the detection threshold.
10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-8.