Electromagnetic fingerprinting method based on frequency domain constellation characteristics

Abstract

The application discloses a kind of electromagnetic fingerprint identification methods based on frequency domain constellation, comprising the following steps: obtaining the near-field electromagnetic radiation signal of chip to be identified, and extract the electromagnetic radiation signal segment of the starting to terminal moment of the function being executed by chip;The electromagnetic radiation signal segment is carried out fast fourier transform, and frequency domain amplitude normalization is carried out, and the frequency domain constellation is constructed using the normalized spectrum feature of the electromagnetic radiation signal segment;The frequency domain constellation is input into the convolutional neural network after training is completed, and the execution function of chip is identified.The application is applied to signal identification field, by the way of frequency domain constellation, the electromagnetic leakage characteristics of chip are displayed, to highlight the unique characteristics of different function electromagnetic leakage, not only have high recognition accuracy, low complexity, but also have unique advantages in anti-gaussian white noise interference.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, specifically an electromagnetic fingerprint recognition method based on frequency domain constellation diagram features. Background Technology

[0002] Side-channel attacks based on chip electromagnetic leakage are a newly emerging physical layer attack method used in recent years to identify, determine the function of, and predict the remaining lifespan of chips or integrated circuits. Due to defects in the chip manufacturing process, there are subtle differences between internal components and bonding wires. Multifunctional chips also exhibit differences in internal conduction circuits and operating frequencies when performing different functions; these differences are collectively referred to as chip electromagnetic fingerprints. A chip's fingerprint characteristics are unique and persistent, and can be considered the chip's "DNA." Chip fingerprints are presented in the form of unintentional electromagnetic leakage. Through advanced electromagnetic measurement techniques and feature extraction methods, it is possible to identify individual chips, their functions, and their aging status.

[0003] Existing theories have proven that unintentional near-field radiation from the surface of MCU chips carries information about the functions being executed. By analyzing and comparing the near-field radiation of ICs, different functions can be distinguished, thus enabling IC function identification. Initially, research involved collecting the radiation field from the MCU surface using a near-field test antenna and applying convolutional neural networks to identify the functions based on the time-frequency maps of the radiation fields. The final function identification accuracy reached 30% without grouping. Subsequently, by converting the time-domain data to transform-domain data and improving the classification method, the accuracy of single-function identification was increased to 37.78%. These research results do not show a high identification rate for the radiation field of integrated circuit surfaces, mainly because a suitable feature extraction method has not been found, resulting in insufficient ability of the algorithm to extract functional features from near-field data.

[0004] To extract more functional features, subsequent research optimized the testing scheme, improving from near-field single-point acquisition to multi-point data acquisition controlled by a robotic arm. Simultaneously, the sampling rate was increased from 100 MSa / s to 1.428 GSa / s, ultimately improving the functional recognition rate for the MCU to 39.24% and for the FPGA to 98.71%. While the FPGA functional recognition rate is satisfactory, improving feature extraction capabilities from a testing perspective is not optimal. This is because it requires more testing time, increasing testing complexity, and handling a larger volume of data, thus increasing computational complexity. The current challenge is to preprocess the data from a single test to highlight more feature information for the classification and recognition algorithm to extract features.

[0005] In the field of radio frequency fingerprinting, differential constellation diagrams are widely used for signal feature extraction. Researchers can distinguish different devices by receiving and demodulating the I and Q signals emitted by radio frequency transmitting devices and presenting them on a CTF (Content-to-Finger Map). Inspired by CTF, we consider displaying the electromagnetic leakage features of chips using a frequency domain constellation diagram (FCTF) to highlight the unique characteristics of electromagnetic leakage for different functions. CNNs have many successful applications in image recognition and classification tasks. Inspired by this, this invention applies CNNs to FCTF classification of electromagnetic leakage data for different chip functions to improve the recognition rate. We also compare the recognition accuracy under different signal-to-noise ratios, demonstrating the advantages of the proposed method in resisting Gaussian white noise. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides an electromagnetic fingerprint recognition method based on frequency domain constellation diagram features. This method displays the electromagnetic leakage characteristics of a chip using a frequency domain constellation diagram, highlighting the unique characteristics of electromagnetic leakage for different functions. It not only has high recognition accuracy and low complexity, but also has unique advantages in resisting Gaussian white noise interference.

[0007] To achieve the above objectives, the present invention provides an electromagnetic fingerprint recognition method based on frequency domain constellation diagram features, comprising the following steps:

[0008] Step 1: Obtain the near-field electromagnetic radiation signal of the chip to be identified, and extract the electromagnetic radiation signal segment from the start to the end of the function being performed by the chip.

[0009] Step 2: Perform a fast Fourier transform on the electromagnetic radiation signal segment and normalize the frequency domain amplitude. Construct a frequency domain constellation diagram using the normalized spectral characteristics of the electromagnetic radiation signal segment.

[0010] Step 3: Input the frequency domain constellation diagram into the trained convolutional neural network to identify the chip's execution functions.

[0011] In one embodiment, step 1 specifically includes:

[0012] Record the near-field electromagnetic radiation signal of the chip to be identified using an oscilloscope;

[0013] IO-Mark technology is used to mark the start and end times of the functions being performed by the chip.

[0014] Based on the start and end times of the chip's function execution, the near-field electromagnetic radiation signal recorded by the oscilloscope is extracted to obtain the electromagnetic radiation signal segment from the start to the end time of the chip's function execution.

[0015] In one embodiment, step 2, which involves constructing a frequency domain constellation diagram using the normalized spectral features of the electromagnetic radiation signal segment, specifically includes:

[0016] The frequency domain constellation diagram is constructed by using the normalized imaginary part of the spectrum in the electromagnetic radiation signal segment as the horizontal axis and the normalized real part of the spectrum in the electromagnetic radiation signal segment as the vertical axis.

[0017] In one embodiment, in step 2, the convolutional neural network has three convolutional layers and one fully connected layer;

[0018] The kernel size of the convolutional layer is 5×5 with a stride of 1 to capture chip fingerprint details in the frequency domain constellation diagram.

[0019] The three convolutional layers have 16, 32, and 64 channels, respectively, to obtain higher-dimensional features;

[0020] The convolutional layer uses rectified linear units as activation functions to reduce the possibility of gradient vanishing and speed up training.

[0021] The fully connected layer performs functional classification via the Softmax activation function.

[0022] In one embodiment, the outputs of the first two convolutional layers in the convolutional neural network are each connected to a 2×2 max pooling layer to downsample the data features and limit the size of the neural network.

[0023] In one embodiment, the process of the convolutional neural network is as follows: the loss function is Softmax loss, and the training algorithm is gradient descent.

[0024] Compared with the prior art, the present invention has the following beneficial technical effects:

[0025] This invention displays the electromagnetic leakage characteristics of chips using a frequency domain constellation diagram to highlight the unique characteristics of electromagnetic leakage for different functions. The accuracy of functional identification of MCU chips can reach over 99%. It not only has high identification accuracy and low complexity, but also has unique advantages in resisting Gaussian white noise interference. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0027] Figure 1This is a flowchart of the electromagnetic fingerprint recognition method based on frequency domain constellation diagram features in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of the time-domain test waveform and amplitude-frequency characteristics of the surface radiation field of the MCU when performing ADC and DAC functions in an embodiment of the present invention, wherein: Figure 2 (a) is a schematic diagram of the time-domain test waveform when the ADC function is executed. Figure 2 (b) is a schematic diagram of the amplitude-frequency response when the ADC function is executed. Figure 2 (c) is a schematic diagram of the time-domain test waveform when the DAC function is executed. Figure 2 (d) is a schematic diagram of the amplitude-frequency characteristics when the DAC function is executed;

[0029] Figure 3 This is the frequency domain constellation diagram of the MCU executing the ADC and DAC in this embodiment of the invention, wherein: Figure 3 (a) Frequency domain constellation diagram when performing ADC function. Figure 3 (b) Frequency domain constellation diagram when performing DAC function;

[0030] Figure 4 This is a schematic diagram of the result of the convolutional neural network in an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram illustrating the variation range of signal points in an embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram of the near-field radiation data of the MCU during the instruction cycle in an embodiment of the present invention;

[0033] Figure 7 This is a schematic diagram of the training results of the convolutional neural network in an embodiment of the present invention, wherein: Figure 7 (a) is a schematic diagram of the training and testing results. Figure 7 (b) is a schematic diagram showing the recognition accuracy of different image qualities;

[0034] Figure 8 This is a schematic diagram illustrating the training results of a convolutional neural network using different learning rates in an embodiment of the present invention, wherein: Figure 8 (a) is a schematic diagram of the training results at a learning rate of 0.001. Figure 8 (b) is a schematic diagram of the training results at a learning rate of 0.01. Figure 8 (c) is a schematic diagram of the training results at a learning rate of 0.1;

[0035] Figure 9 This is a schematic diagram comparing the recognition accuracy in an embodiment of the present invention;

[0036] Figure 10 A schematic diagram of the confusion matrix in an embodiment of the present invention, wherein, Figure 10 (b) is a schematic diagram of the confusion matrix at a signal-to-noise ratio of -10 dB. Figure 10 (b) is a schematic diagram of the confusion matrix at a signal-to-noise ratio of -15dB.

[0037] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0039] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0040] This embodiment discloses an electromagnetic fingerprint recognition method based on frequency domain constellation diagram features. By processing data from a single test of a chip (integrated circuit), the features are displayed in the form of a frequency domain constellation diagram. Simultaneously, by combining a convolutional neural network classification method, the method effectively improves the accuracy of functional recognition of the chip surface radiation field. (Reference) Figure 1 The electromagnetic fingerprint recognition method based on frequency domain constellation diagram features in this embodiment specifically includes the following steps:

[0041] Step 1: Acquire the near-field electromagnetic radiation signal of the chip to be identified, and extract the electromagnetic radiation signal segment from the start to the end of the function being performed by the chip. The specific process is as follows:

[0042] Record the near-field electromagnetic radiation signal of the chip to be identified using an oscilloscope;

[0043] IO-Mark technology is used to mark the start and end times of the functions being performed by the chip.

[0044] Based on the start and end times of the chip's function execution, the near-field electromagnetic radiation signal recorded by the oscilloscope is extracted to obtain the electromagnetic radiation signal segment from the start to the end time of the chip's function execution.

[0045] Step 2: Perform a Fast Fourier Transform on the electromagnetic radiation signal segment and normalize its frequency domain amplitude. Construct a frequency domain constellation diagram using the normalized spectral characteristics of the electromagnetic radiation signal segment, specifically:

[0046] A frequency domain constellation diagram is constructed using the imaginary part of the normalized spectrum in the electromagnetic radiation signal segment as the horizontal axis and the real part of the normalized spectrum in the electromagnetic radiation signal segment as the vertical axis.

[0047] Step 3: Input the frequency domain constellation diagram into the trained convolutional neural network to identify the chip's execution functions.

[0048] When an MCU (Microcontroller Unit) chip is working, its internal CPU controls the operating frequency through frequency division or combination, and controls the on / off state of different functional peripheral circuits via instructions. Some functions of the MCU operate at different main frequencies, resulting in significantly different amplitude-frequency characteristics in the near-field received signal frequency domain, which can be easily distinguished using common data classification algorithms. However, more generally, some MCU functions execute at the same main frequency, making it difficult to distinguish them solely from a frequency domain perspective. In this case, the signal coupling path can serve as a distinguishing factor. Even when the main frequency is the same when the chip executes different functions, the internal signal paths will inevitably differ. Due to differences in trace location, length, and time delay, different connection paths will cause the radiated electromagnetic waves coupled to the near-field receiver to exhibit different amplitude and phase characteristics. These characteristics can serve as a basis for distinguishing different functions.

[0049] For example Figure 2 This describes the time-domain waveform and amplitude-frequency characteristics of the surface radiation field when the MCU performs ADC and DAC functions. Figure 2 (b) Figure 2 As can be seen from the amplitude-frequency characteristic distribution in (d), the bottom noise distribution of the two signals is almost the same, with only a few high-amplitude frequency points representing the chip's operating frequency and its higher harmonics. The main difference between the two signals in the frequency domain lies in these high-amplitude frequency points. Directly using the frequency domain data of the test signal for classification and recognition is inaccurate because a large amount of bottom spurious noise "dilutes" the feature information carried by the main frequency and its harmonics. Conventional recognition algorithms allocate a large amount of computing power to processing spurious data when performing classification tasks, resulting in low algorithm efficiency and low recognition rate. Filtering out redundancy before classification is also not advisable, because although the main functional features are hidden in a few main frequencies and their harmonics, other frequency components also contain a small amount of features. Although these frequencies contain few feature components, they account for the vast majority of the frequency points within the passband, so the total feature content cannot be ignored and cannot be simply filtered out. Therefore, the best solution to improve feature extraction capability is to process the original data before classification to highlight the features carried by the high-amplitude frequency points and weaken the features of the bottom redundant frequency points.

[0050] Constellation trajectory diagrams have wide applications in the field of radio frequency fingerprint recognition. The principle is to demodulate the received signal and then perform differential operations on the baseband signal to convert the IQ signal into a stable graph, displaying the amplitude and phase information of the signals transmitted by different devices on a single graph. Due to subtle differences between individual devices, they will exhibit different distribution patterns on the constellation trajectory diagram. Different individuals can be distinguished using classification and recognition methods from the field of image recognition. IC near-field radiation waves are unmodulated broadband electromagnetic waves, significantly different from intentionally transmitted radio frequency signals. Their broadband characteristics make it impossible to extract amplitude and phase information from the time domain to draw a constellation trajectory diagram. Therefore, this embodiment adopts a feature enhancement scheme based on frequency domain constellation diagrams. Analogous to the method of drawing time-domain constellation trajectory diagrams, the horizontal axis represents the normalized imaginary part of the spectrum, and the vertical axis represents the normalized real part of the spectrum. The single-cycle amplitude and phase data of the radiated electromagnetic wave are extracted and drawn on the frequency domain constellation diagram. A single-cycle radiation signal for each function can be used to draw a constellation diagram, for example... Figure 3 This refers to the frequency domain constellation diagram corresponding to the ADC and DAC executed by the MCU.

[0051] This embodiment converts radiation data into a frequency domain constellation diagram. Frequency points with larger amplitudes are scattered on the periphery, making them easier to extract features from an image recognition perspective. Conversely, frequency points with smaller amplitudes cluster near the origin, making it difficult to extract the feature information carried by individual frequency points. This achieves the goal of highlighting the features of high-value frequencies while weakening the features of low-value frequencies.

[0052] Since the frequency domain constellation diagram is a two-dimensional image, a mature CNN model can be used for classification and recognition. This embodiment designs a convolutional neural network with three convolutional layers and one fully connected layer, namely… Figure 4 As shown, the input image size of the convolutional neural network (CNN) is limited to 227×227×3, the kernel size of the convolutional layers is 5×5, and the stride is 1 to capture chip fingerprint details in the frequency domain constellation map. The number of channels in the three convolutional layers are 16, 32, and 64, respectively, to obtain higher-dimensional features. Simultaneously, rectified linear units (RCUs) are used as activation functions in the convolutional layers to reduce the possibility of gradient vanishing and accelerate training. Furthermore, 2×2 max-pooling layers are connected to the outputs of the first two convolutional layers in the CNN to downsample the data features and limit the size of the neural network. Finally, a fully connected layer and a softmax activation function are used to perform functional classification. The CNN process uses softmax loss as the loss function and gradient descent as the training algorithm. Specifically, the parameters of each layer of the CNN are as follows:

[0053] The first convolutional layer has 5×5×1×16 weights and 16 biases, with 416 parameters to be trained. After a 2×2 max pooling layer, the output is 112×112.

[0054] The second convolutional layer has 5×5×16×32 weights and 32 biases, with 12832 parameters to be trained. After a 2×2 max pooling layer, the output is 54×54.

[0055] The third convolutional layer has 5×5×32×64 weights and 64 biases, with 51264 parameters to be trained. After passing through a 2×2 max pooling layer, the output is 25×25.

[0056] The fully connected layer has: since the final image size is 25×25 and the number of channels is 64, there are a total of 25×25×64 weight parameters and 1 bias; when there are L classification targets, the number of trainable parameters is L×40001.

[0057] The total weights and bias parameters of the network are: 416 + 12832 + 51264 + L × 40001 = 64512 + L × 40001.

[0058] In this embodiment, the electromagnetic leakage characteristics of the chip are displayed using a frequency domain constellation diagram to highlight the unique characteristics of electromagnetic leakage for different functions, and it also has unique advantages in resisting Gaussian white noise interference. Since the near-field radiation signal has a finite bandwidth, Gaussian white noise is band-limited, and its power spectral density... Represented as:

[0059]

[0060] in, The constant represents the power density per unit frequency within the band limit. For frequency, and The start frequency and the cutoff frequency

[0061] According to the formula for calculating signal-to-noise ratio (SNR), we have:

[0062]

[0063] in, For noise power, For signal power For time-domain signals, and These are the start and stop times of the signal, respectively.

[0064] According to Bashwa's equation, signal power It can also be expressed as:

[0065]

[0066] Among them, The spectrum of the signal;

[0067] Substituting the signal power from equation (3) and the noise power from equation (1) into equation (2) and transforming them, we can obtain the power spectral density of the band-limited noise as follows:

[0068]

[0069] The power spectral density of the signal It can also be expressed as:

[0070]

[0071] Substituting the signal power spectral density from equation (5) into equation (4) yields:

[0072]

[0073] in, It can be considered as the average power density of the signal.

[0074] From the perspective of SNR, when SNR is set, The noise is constant and evenly distributed throughout the passband. Since the signal energy is relatively concentrated within the passband, the noise has a smaller impact on frequencies with large amplitudes and a larger impact on frequencies with small amplitudes. To accurately quantify the impact of noise on signals of different amplitudes, this embodiment defines the power density ratio as:

[0075]

[0076] in, express The power density at point M is The logarithm of the ratio of the power density at a given point to the average power density of the signal;

[0077] Given SNR, frequency The ratio of signal power density to noise power density at that location is

[0078]

[0079] i.e., frequency point The actual signal-to-noise ratio of the signal affected by noise is Since the electromagnetic leakage energy of the chip is mainly concentrated at the operating frequency and its harmonics, the M-values ​​at these energy-concentrated frequencies are generally greater than 40 dB, while the M-values ​​at spurious frequencies at the bottom of the spectrum are generally less than -10 dB. Therefore, the actual signal-to-noise ratio of these high-amplitude frequencies to Gaussian white noise is greater than... The signal-to-noise ratio of spurious frequencies to Gaussian white noise is less than that of spurious frequencies. .

[0080] From the perspective of frequency domain constellation diagrams, let the distance from the signal point to the origin be F (i.e., frequency domain amplitude). When the SNR is given, the signal point may change to... Figure 5 The radius of the circle can be calculated based on the signal-to-noise ratio at any position within the circle shown.

[0081]

[0082] Table 1 lists the radius of variation of different signal amplitudes under different signal-to-noise ratios.

[0083] Table 1

[0084]

[0085] As shown in Table 1, under low signal-to-noise ratio (SNR) conditions, large-amplitude signals carrying more feature information are less affected by noise, while small-amplitude spurious waves are more affected. Small-amplitude spurious waves are numerous and exhibit a clustering effect in the frequency domain constellation diagram. Although the position of individual points in the constellation diagram changes significantly after noise is added, their statistical characteristics remain largely concentrated near the origin, thus not significantly affecting the feature classification of the frequency domain constellation diagram. Large-amplitude frequency points are sparsely distributed on the periphery of the frequency domain constellation diagram, and large blank areas often exist around individual points. Even with a certain degree of interference, their range of variation, without overlapping with surrounding points, still possesses unique characteristics and can be used as feature points to distinguish different classes.

[0086] From the perspective of classification and recognition methods, the change in position of feature points far from the center region on the frequency domain constellation diagram is equivalent to the image pixels being scrambled. Convolutional neural networks have a unique advantage in recognizing images with scrambled pixels:

[0087] 1. The convolutional layers of a convolutional neural network extract local features of an image through local receptive fields, and can still detect features such as corner points even if the pixels are scrambled;

[0088] 2. Pooling layers further enhance the robustness of convolutional neural networks to pixel position changes by reducing the dimension of feature maps. Pooling operations ignore the precise position of pixels to a certain extent and focus on the statistical properties of features, enabling convolutional neural networks to identify important features even when pixels are shuffled.

[0089] 3. The fully connected layer combines and abstracts the features extracted earlier to form higher-level feature representations. These higher-level features are usually not sensitive to the specific location of pixels, but rather focus on the overall structure and semantic information of the image; therefore, even if the pixels are scrambled, the fully connected layer can still perform classification based on these abstract characteristics.

[0090] 4. Convolutional neural networks represent pixel features in a distributed form through multiple layers of convolution and pooling operations. This means that feature information is distributed across multiple neurons and multiple layers, rather than concentrated in a specific location. Therefore, even if pixels are shuffled, the distributed representation of features can still retain enough information for classification.

[0091] In summary, Gaussian white noise has relatively little impact on the main features of the frequency domain constellation diagram. Furthermore, the convolutional neural network classification method can also reduce the impact of perturbations on recognition accuracy. Therefore, the electromagnetic fingerprint recognition method based on frequency domain constellation diagram features in this embodiment has a unique advantage in its resistance to Gaussian white noise.

[0092] The following examples provide further illustration of the electromagnetic fingerprint recognition method based on frequency domain constellation graph features in this embodiment.

[0093] The experimental system uses a stripline cell as the core device for chip radiated emission testing. The stripline cell effectively shields against external interference, resulting in more accurate measurement results. Furthermore, the closer proximity of the chip board to the test object facilitates the capture of subtle near-field radiation characteristics. The microcontroller is controlled by instructions to perform different functions. The stripline cell receives surface radiation waves from the MCU, and the radiation data is read and recorded using an oscilloscope. The oscilloscope has a sampling rate of 1 GSa / s and dual input channels: one channel displays the surface radiation data received by the stripline cell, and the other channel displays the execution flags output by the microcontroller pins.

[0094] During the experiment, the STM32F103ZET6 microcontroller was used and placed on one side of the PCB board, while the other peripheral components were placed on the other side. Fourteen functions of the MCU were selected, and the surface radiation field during the execution of each function was tested individually. The functional information executed by the MCU is shown in Table 2.

[0095] Table 2

[0096]

[0097] In Table 2: ADC represents analog-to-digital conversion, LED represents pin output signal on / off, DAC represents digital-to-analog conversion, PWM represents pulse width modulation, DMA represents direct memory access, IWDG represents independent watchdog timer, WWDG represents window watchdog timer, MALLOC represents memory management experiment, USART represents serial communication, TIMER represents timer, TSENSOR represents internal temperature sensor, PWM+DAC represents pulse width modulation and digital-to-analog conversion, PWM+ADC represents pulse width modulation and analog-to-digital conversion, and PWM+DAC+ADC represents pulse width modulation, digital-to-analog conversion, and analog-to-digital conversion.

[0098] All measurements during the experiment were performed according to IEC 61967-8. The electromagnetic field radiated by the MCU under test was received through a stripline cell and transmitted to the oscilloscope. Simultaneously, IOMARK technology was used to mark the commands. For example... Figure 6 As shown, when the instruction begins execution, the designated I / O port of the MCU starts outputting a high level, and when the instruction execution ends, the I / O port outputs a low level. Each execution of an instruction is called an instruction cycle, and the oscilloscope records the near-field radiation data within that cycle.

[0099] In addition, to reduce the impact of peripheral components on functional classification, five PCB boards were fabricated in the experiment. Each PCB board was used to test 14 functions, and 19-22 functional cycles were saved for each test, resulting in a total of 1385 cycle data points for different instructions. To increase the dataset and assess noise immunity, nine sets of Gaussian white noise at different levels were added to the test data, with a signal-to-noise ratio ranging from 0 to 32 dB, in 4 dB increments. After adding Gaussian white noise, the original test data and the original data together formed 13850 sets of instruction cycle data, and 13850 frequency domain constellation diagrams were plotted.

[0100] The convolutional neural network was trained and tested using the prepared data. During the initial training, the batch size was set to 64, the maximum epoch was 30, and the initial learning rate was 0.001. Figure 7 (a) shows the training and testing results. It can be seen that after about 20 epochs of training and testing, the accuracy reached 99.9%, and there was no overfitting.

[0101] The pixel size of the frequency domain constellation map affects recognition accuracy and training time. A higher pixel size reduces the probability of different frequency points falling into the same pixel, allowing for the display of more frequency domain feature details; however, increased pixel size prolongs network training time. Conversely, smaller pixel size reduces computational complexity and training time, but it also blurs frequency domain features. Figure 7 (b) shows the recognition accuracy for different image qualities. It can be seen that 227×227 pixels achieves the highest recognition accuracy (99.99%), while 57×57 and 113×113 pixels also reach over 99%. When the pixel count is below 29×29, the recognition accuracy drops below 90%. Table 3 shows the training time for different image qualities to achieve the highest recognition accuracy. It can be seen that the overall complexity of the network is not high; the training time for 13850 samples with 227×227 pixels is only 82 minutes. Smaller image pixels result in shorter training time, but the features become more difficult to distinguish, requiring more epochs to achieve optimal training accuracy.

[0102] Table 3

[0103]

[0104] The efficiency and accuracy of convolutional neural networks are affected by batch size, maximum epochs, and initial learning rate. Larger batches can accelerate the training process but may increase memory usage; smaller batches, while reducing memory usage, will prolong training time. Increasing the number of epochs can improve the network's training accuracy but may also lead to overfitting. Larger learning rates can speed up training but may cause the network to oscillate around the optimal solution. In further optimizing the method of this embodiment, 13850 frequency domain constellation maps of size 57×57 were used for training. Performance was evaluated at different batch sizes (from 32 to 1024) and different maximum epochs (from 5 to 40) with initial learning rates of 0.001, 0.01, and 0.1. The results are as follows... Figure 8 As shown.

[0105] from Figure 8 It can be seen that the lower the learning rate, the more stable the training process. When the initial learning rate is 0.001, the network trains faster and more accurately. System performance is better when the batch size is 64 or 128; larger batches require more epochs to achieve high accuracy, while smaller batches require more computation time. Furthermore, even with longer computation time, smaller batches lead to a decrease in accuracy when the initial learning rates are 0.01 and 0.1.

[0106] Furthermore, the classification accuracy of the method in this embodiment was evaluated within a signal-to-noise ratio range of {-20, -15, -10, -5, 0} dB at different signal-to-noise ratios. The frequency domain constellation map had a pixel size of 227×227, a learning rate of 0.001, a batch size of 32, and a maximum epoch of 40.

[0107] Figure 9 The electromagnetic fingerprint recognition method based on frequency domain constellation diagram features (PCTF-CNN) in this embodiment was compared with existing one-dimensional CNN classification methods based on time domain data (TIME-CNN), one-dimensional CNN methods based on frequency domain data (PREO-CNN), and time domain data classification methods based on the nearest neighbor algorithm (KNN). All evaluations used the same dataset. The one-dimensional convolutional neural network employed a three-layer convolutional structure with a kernel size of 5×1 and channel numbers of 16, 32, and 64, respectively. The confusion matrix obtained by the method in this embodiment for 14 MCU function recognition problems at SNRs of -10dB and -15dB is shown below. Figure 10 As shown. When the signal-to-noise ratio is greater than -5dB, the accuracy of the method in this embodiment is higher than 95%, and when it is above 0dB, the accuracy is greater than 99.9%.

[0108] Due to the feature representation of the frequency domain constellation graph and the classification capability of the convolutional neural network, the method in this embodiment achieves a significant performance improvement in low signal-to-noise ratio (SNR) scenarios. It can achieve a recognition accuracy of 64.98% for 14 functions at an SNR of -20dB, while the accuracy of a one-dimensional CNN based on time-domain data is 26.6%, a one-dimensional CNN based on frequency-domain data is 29.9%, and the KNN recognition accuracy is only 7.8%. In high SNR scenarios, the electromagnetic fingerprint recognition method based on frequency-domain constellation graph features in this embodiment has slightly higher recognition accuracy than the one-dimensional CNN based on both time-domain and frequency-domain data.

[0109] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An electromagnetic fingerprint recognition method based on frequency domain constellation diagram features, characterized in that, Includes the following steps: Step 1: Acquire the near-field electromagnetic radiation signal of the chip to be identified, and extract the electromagnetic radiation signal segment from the start to the end of the function being performed by the chip, specifically including: Record the near-field electromagnetic radiation signal of the chip to be identified using an oscilloscope; IO-Mark technology is used to mark the start and end times of the functions being performed by the chip. Based on the start and end times of the chip's function execution, the near-field electromagnetic radiation signal recorded by the oscilloscope is extracted to obtain the electromagnetic radiation signal segment from the start to the end time of the chip's function execution. Step 2: Perform a Fast Fourier Transform on the electromagnetic radiation signal segment and normalize its frequency domain amplitude. Construct a frequency domain constellation diagram using the normalized spectral characteristics of the electromagnetic radiation signal segment. Specifically: The frequency domain constellation diagram is constructed by using the normalized imaginary part of the spectrum in the electromagnetic radiation signal segment as the horizontal axis and the normalized real part of the spectrum in the electromagnetic radiation signal segment as the vertical axis. Step 3: Input the frequency domain constellation diagram into the trained convolutional neural network to identify the chip's execution functions.

2. The electromagnetic fingerprint recognition method based on frequency domain constellation diagram features according to claim 1, characterized in that, In step 2, the convolutional neural network has three convolutional layers and one fully connected layer; The kernel size of the convolutional layer is 5×5 with a stride of 1 to capture chip fingerprint details in the frequency domain constellation diagram. The three convolutional layers have 16, 32, and 64 channels, respectively, to obtain higher-dimensional features; The convolutional layer uses rectified linear units as activation functions to reduce the possibility of gradient vanishing and speed up training. The fully connected layer performs functional classification via the Softmax activation function.

3. The electromagnetic fingerprint recognition method based on frequency domain constellation diagram features according to claim 2, characterized in that, The outputs of the first two convolutional layers in the convolutional neural network are connected to 2×2 max pooling layers to downsample the data features and limit the size of the neural network.

4. The electromagnetic fingerprint recognition method based on frequency domain constellation diagram features according to claim 2, characterized in that, The training process of the convolutional neural network uses Softmax loss as the loss function and gradient descent as the training algorithm.

Images