An electroencephalogram deep feature information evaluation method based on frequency domain reconstruction

Abstract

The application provides a brain electrical deep feature information evaluation method based on frequency domain reconstruction. The brain electrical deep compressed features in a communication channel are acquired, and an identity inference network is input to extract a subject identity correlation index. A frequency domain feature generator is used to reconstruct an initial brain electrical power spectrum density atlas, and a reconstruction loss and a fidelity index are obtained by comparing the atlas with a real atlas. For a cross-domain scene, a semantic constraint loss between the atlas and a real label is calculated, and is combined with the reconstruction loss to optimize the generator to output a final atlas. The initial and final reconstructed atlases are respectively input into a target classification network together with the real atlas, and a target state semantic consistency rate of the basic and cross-domain scenes is calculated. Finally, the consistency rate, the identity correlation index and the fidelity index are comprehensively weighted to output a comprehensive information evaluation result in the double scenes. The method breaks away from the white box strong hypothesis, and provides an objective and comprehensive deep feature information comprehensive evaluation benchmark for an end-to-cloud collaborative system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of EEG signal processing, artificial intelligence and information security. Specifically, it is a method for evaluating deep EEG feature information based on frequency domain reconstruction. Background Technology

[0002] Automatic classification and recognition of abnormal states based on electroencephalogram (EEG) signals has become a research area that has received widespread attention in recent years at the intersection of brain cognitive science and artificial intelligence. With the continuous advancement of wearable sensor technology, in order to alleviate the computational pressure on edge devices and reduce the direct exposure of raw sensitive signals, current EEG classification systems generally adopt an "edge-cloud collaborative" architecture: multi-channel EEG waveforms are acquired by edge devices, and through feature encoding networks and dimensionality reduction mapping layers, they are nonlinearly compressed into low-dimensional abstract deep compression features of EEG. Subsequently, only these features are transmitted across physical channels to the cloud to perform the final target state assessment.

[0003] In this paradigm, both academia and industry have generally fallen into a "delusion of inherent security." The mainstream view assumes that edge-side deep networks act as a severe information bottleneck in dimensionality reduction operations such as feature extraction and global pooling, irreversibly truncating the absolute temporal phase of the underlying signal and forcibly stripping away redundant physiological noise. Therefore, these highly condensed deep feature vectors are widely considered to have naturally achieved data anonymization. This assumption that "feature compression is equivalent to privacy protection" leads to a severe underestimation of the risk of real underlying data exposure when existing systems are subjected to eavesdropping on open communication channels.

[0004] However, most existing feature evaluation and privacy verification methods rely on unrealistic strong capability assumptions: either requiring the evaluator to have complete white-box access to the edge extractor (to obtain the complete network architecture and parameter weights), or allowing unrestricted oracle queries on the target cloud service to train the proxy model. This stringent setting, detached from actual communication links, artificially raises the risk verification threshold, failing to objectively and accurately quantify the true vulnerability of edge-cloud systems in black-box interception scenarios. Furthermore, existing evaluation metrics are often too simplistic, mostly limited to shallow statistical errors within the same domain dataset, severely lacking consideration for the semantic meaning of the target state in cross-domain scenarios caused by heterogeneous devices. This also prevents existing methods from exploring whether the high-fidelity reconstruction of the physical form of deep features after interception necessarily involves the leakage of related deep target state semantics.

[0005] Therefore, designing a more effective method for evaluating deep EEG features that does not rely on strong white-box assumptions and has outstanding effects on multidimensional leakage assessment, accurate quantification of identity tracing and semantic consistency in cross-domain scenarios has become an urgent technical problem to be solved. Summary of the Invention

[0006] In view of this, and addressing the problems of existing feature evaluation methods being dimensionally limited, overly reliant on strong white-box assumptions, and lacking a true measure of data exposure risk in cross-domain scenarios, this invention proposes a method for evaluating deep EEG feature information based on frequency domain reconstruction. To achieve the above objectives, this invention provides the following technical solution:

[0007] A method for evaluating deep EEG feature information based on frequency domain reconstruction includes the following steps:

[0008] Step S1: Obtain the deep compression features of EEG transmitted in the communication channel and the corresponding real EEG power spectral density map. The deep compression features of EEG are generated by the end device inputting the original EEG signal into the feature coding network to extract the spatiotemporal high-dimensional features, and then compressing the spatiotemporal high-dimensional features into a low-dimensional vector space through the dimension reduction mapping layer.

[0009] Step S2: Input the deep compression features of EEG into the pre-trained identity inference network, calculate the probability of the deep compression features of EEG belonging to the identity of each candidate subject, and determine the identity prediction accuracy based on the probability ranking, and use it as the identity tracing evaluation index.

[0010] Step S3: Input the deep compression features of the EEG into the frequency domain feature generator to reconstruct and generate the corresponding EEG power spectral density map;

[0011] Step S4: Based on the reconstructed EEG power spectral density map and the real EEG power spectral density map, calculate the mean square error and cosine similarity as the map reconstruction fidelity index, and construct a reconstruction loss function based on the mean square error and the cosine similarity.

[0012] Step S5: In cross-domain scenarios where there are differences in data features between the source data and the auxiliary dataset, the alternative classification network is pre-trained using the auxiliary dataset. After fixing its network parameters, the reconstructed EEG power spectral density map is input into the alternative classification network, and the semantic constraint loss between the network output and the real target state label is calculated.

[0013] Step S6: Combine the reconstruction loss function with the semantic constraint loss to construct a target total loss function; update the network parameters of the frequency domain feature generator based on the target total loss function to reconstruct and generate the final EEG power spectral density map;

[0014] Step S7: Input the EEG power spectral density maps generated in Step S3 and Step S6, along with the actual EEG power spectral density maps, into the target classification network with fixed parameters, and calculate the semantic consistency rate of the target state.

[0015] Step S8: Based on the consistency rate, the identity tracing evaluation index, and the map reconstruction fidelity index, output the comprehensive evaluation result of the feature information of the deep compression features of the EEG.

[0016] Further, step S2 includes the following steps:

[0017] Step 1: Based on the raw EEG signal features with subject labels in the auxiliary dataset, the identity inference network is pre-trained;

[0018] Step 2: Input the deep EEG compression features into the pre-trained identity inference network, and calculate the probability that the deep EEG compression features belong to the identities of each candidate subject. The specific calculation formula is as follows:

[0019] ,

[0020] in, Indicates belonging to the first The probability of each candidate's identity. This indicates the deep compression feature of the electroencephalogram (EEG). The output of the identity inference network represents the first... The log odds of each candidate subject Indicates the first The log odds of a candidate subject;

[0021] Step 3: Sort the identity probabilities in descending order of their numerical values, and select the top-ranked probabilities. Candidate subjects are selected to form a prediction set; if the real subject corresponding to the deep compression feature of the EEG exists in the prediction set, it is determined as a single prediction hit, and the Top-K identity prediction accuracy is calculated based on the number of hits, which is used as an evaluation indicator for identity tracing. It is a positive integer.

[0022] Furthermore, in step S3, the aforementioned deep compression features of EEG are used as input to construct a frequency domain feature generator network containing non-negative activation functions. The generator is used to perform dimensionality analysis and expansion on the input compression features to generate a non-negative numerical matrix that matches the total number of electrode channels and the total number of frequency bands in the real EEG signal, which serves as the initial reconstructed EEG power spectral density map.

[0023] Further, step S5 includes the following steps:

[0024] Step 1: Pre-train an alternative classification network using the auxiliary dataset. Take the real EEG power spectral density map in the auxiliary dataset as input and the real target state label as reference. Iteratively optimize the network parameters of the alternative classification network by minimizing the classification error.

[0025] Step 2: Fix the network parameters of the alternative classification network, input the reconstructed EEG power spectral density map into the alternative classification network, and calculate the semantic constraint loss between the network output and the real target state label. The specific calculation formula is as follows:

[0026] ,

[0027] in, Represents semantic constraint loss. This represents the EEG power spectral density map generated by the reconstruction. Indicates the actual target state label. This represents the total number of categories of the target state. Indicates category index, The label representing the true target state is a category. One-hot encoding value at time, This indicates that the reconstructed EEG power spectral density map belongs to the category. The semantic probability.

[0028] Further, step S6 includes the following steps:

[0029] Step 1: Keep the network parameters of the alternative classification network fixed, calculate the partial derivative of the semantic constraint loss with respect to the reconstructed EEG power spectral density map, and backpropagate the partial derivative as a feedback gradient to the frequency domain feature generator based on the chain rule.

[0030] Step 2: The reconstruction loss function and the semantic constraint loss are weighted and jointly constructed to form a target total loss function. On the auxiliary dataset, the network parameters of the frequency domain feature generator are iteratively updated by minimizing this target total loss function. The final EEG power spectral density map is then reconstructed using the updated frequency domain feature generator. The specific formula for calculating the target total loss function is as follows:

[0031] ,

[0032] in, This represents minimizing the network parameters of the frequency domain feature generator. This represents the network parameters of the frequency domain feature generator. Represents the mathematical expectation. Represents the auxiliary dataset, Represents the reconstruction loss function. This represents the frequency domain feature generator function. This indicates the deep compression feature of the electroencephalogram (EEG). This represents the true power spectral density map of brainwaves. This represents the preset semantic constraint loss weight coefficients. Represents semantic constraint loss. This indicates the actual target status label.

[0033] Further, step S7 includes the following steps:

[0034] Step 1: Input the initial EEG power spectral density map generated in step S3 and the corresponding real EEG power spectral density map into the target classification network with fixed parameters for prediction, and count the proportion of samples that output the same target state prediction results, and use it as the basic scene semantic consistency rate.

[0035] Step 2: Input the final EEG power spectral density map generated in step S6 and the corresponding real EEG power spectral density map into the target classification network with fixed parameters for prediction, and count the proportion of samples that output the same target state prediction results, and use it as the cross-domain scene semantic consistency rate.

[0036] Further, step S8 includes the following steps:

[0037] Step 1: Extract the quantitative values ​​of the identity tracing evaluation index, the quantitative values ​​of the graph reconstruction fidelity index, and the calculated basic scene semantic consistency rate and cross-domain scene semantic consistency rate;

[0038] Step 2: Assign preset risk assessment weight coefficients to identity tracing, graph reconstruction fidelity, and semantic consistency; substitute the basic scenario semantic consistency rate and the cross-domain scenario semantic consistency rate into the linear weighted comprehensive evaluation model to calculate the basic scenario comprehensive information evaluation score and the cross-domain scenario comprehensive information evaluation score. The specific calculation formula is as follows:

[0039] ,

[0040] ,

[0041] in, This indicates the comprehensive information assessment score for the basic scenario. This indicates the comprehensive information evaluation score for cross-domain scenarios. , and These represent the preset risk assessment weight coefficients corresponding to the identity tracing assessment index, the graph reconstruction fidelity index, and the semantic consistency rate, respectively. This represents the quantitative value of the identity tracing and evaluation index. This represents the quantitative value of the fidelity index for the reconstructed map. This represents the semantic consistency rate of the basic scenario. This represents the semantic consistency rate of the cross-domain scenario;

[0042] Step 3: Compare the comprehensive information evaluation score of the basic scenario and the comprehensive information evaluation score of the cross-domain scenario with the preset information retention threshold, and output the comprehensive information evaluation results of the deep compression features of EEG in the basic scenario and the cross-domain scenario, respectively.

[0043] Beneficial effects: This invention proposes a method for evaluating deep EEG feature information based on frequency domain reconstruction, aiming to solve the problems of existing feature evaluation methods having a single dimension, over-reliance on strong white-box assumptions, and lack of real measurement of data exposure risk in cross-domain scenarios.

[0044] By adaptively shifting the reconstruction perspective to the frequency domain (EEG power spectral density map), this method effectively overcomes the temporal phase misalignment problem caused by feature pooling compression. It only requires capturing features in the communication channel to effectively trace the subject's identity and reconstruct the underlying physical morphology map. Furthermore, by combining a cross-domain substitution classification network with a gradient backpropagation strategy, the evaluation framework is endowed with the ability to deeply extract the semantics of the target state. This method constructs a multi-dimensional quantitative standard covering identity tracing, map fidelity, and semantic consistency. It not only effectively overcomes the cross-domain defense barriers caused by heterogeneous data and accurately quantifies the perception and distortion costs in cross-domain scenarios, but also provides an objective and comprehensive evaluation benchmark for deep feature information in edge-cloud collaborative systems. Attached Figure Description

[0045] Figure 1 A flowchart illustrating the steps of a method for evaluating deep EEG feature information based on frequency domain reconstruction, provided in an embodiment of the present invention;

[0046] Figure 2 This is an evaluation curve showing the change in identity prediction accuracy with the number of network iterations in an embodiment of the present invention.

[0047] Figure 3 This is a comparison of the approximation fidelity between the actual EEG power spectral density map and the reconstructed map in each frequency band of a single channel according to an embodiment of the present invention.

[0048] Figure 4 This is a visualization comparison of clustered scatter points of EEG features in the network feature space according to an embodiment of the present invention;

[0049] Figure 5 This is a bar chart comparing the semantic consistency rate of the target state in the basic scenario and the cross-domain scenario of this invention. Detailed Implementation

[0050] To make the technical solution and effects of the present invention clearer, the following description is provided in conjunction with the accompanying drawings and embodiments:

[0051] like Figure 1 As shown, the method for evaluating deep EEG feature information based on frequency domain reconstruction according to the present invention includes the following steps:

[0052] The first step (corresponding to S1) is to obtain the deep compression features of the EEG transmitted in the communication channel and the corresponding real EEG power spectral density map. Specifically, the deep compression features of the EEG are generated by the edge device inputting the original EEG signal into a feature coding network to extract high-dimensional spatiotemporal features, and then compressing the high-dimensional spatiotemporal features into a low-dimensional vector space through a dimensionality reduction mapping layer before outputting them.

[0053] The second step (corresponding to S2) involves inputting the deep EEG compression features into a pre-trained identity inference network, calculating the probability that the deep EEG compression features belong to the identities of each candidate subject, and determining the identity prediction accuracy based on the probability ranking. This accuracy is then used as an identity tracing evaluation indicator. The specific steps are as follows:

[0054] The identity inference network is pre-trained based on the raw EEG signal features with subject labels from the auxiliary dataset. The deep compression features of the EEG are then input into the pre-trained identity inference network to calculate the probability that the deep compression features belong to each candidate subject's identity. The specific calculation formula is as follows:

[0055] ,

[0056] in, Indicates belonging to the first The probability of each candidate's identity. This indicates the deep compression feature of the electroencephalogram (EEG). The output of the identity inference network represents the first... The log odds of each candidate subject Indicates the first The log odds of each candidate subject.

[0057] Sort the identity probabilities in descending order of their numerical values, and select the top-ranked probabilities. Candidate subjects are selected to form a prediction set; if the real subject corresponding to the deep compression feature of the EEG exists in the prediction set, it is determined as a single prediction hit, and the Top-K identity prediction accuracy is calculated based on the number of hits, which is used as an evaluation indicator for identity tracing. It is a positive integer.

[0058] like Figure 2 As shown, Figure 2This is an evaluation curve showing how identity prediction accuracy changes with the number of network iterations. The graph visually demonstrates that as the number of iterations increases, the identity prediction accuracy of Top-1 and Top-5 significantly increases and tends to stabilize and converge, proving that traceable individual identity fingerprints still remain in the deep compressed features.

[0059] The third step (corresponding to S3) involves inputting the deep EEG compressed features into the frequency domain feature generator to reconstruct and generate the corresponding EEG power spectral density map. The specific steps are as follows:

[0060] Using the aforementioned deep compression features of EEG as input, a frequency domain feature generator network containing non-negative activation functions is constructed. The generator performs dimensionality analysis and expansion on the input compression features to generate a non-negative numerical matrix that matches the total number of electrode channels and the total number of frequency bands in the real EEG signal, which serves as the initial reconstructed EEG power spectral density map.

[0061] Step 4 (corresponding to S4): Based on the reconstructed EEG power spectral density map and the real EEG power spectral density map, calculate the mean square error and cosine similarity as indicators of map reconstruction fidelity, and construct a reconstruction loss function based on the mean square error and the cosine similarity.

[0062] like Figure 3 As shown, Figure 3 This is a comparison of the fidelity of the actual EEG power spectral density map and the reconstructed map across different frequency bands in a single channel. Figure 3 It can be seen that the reconstructed spectrum represented by the dashed line achieves a high degree of waveform fitting with the ground truth spectrum represented by the solid line in the main energy frequency band, reflecting the high-fidelity restoration capability of this method at the physical feature level.

[0063] Step 5 (corresponding to S5): In cross-domain scenarios where there are differences in data features between the source data and the auxiliary dataset, a substitute classification network is pre-trained using the auxiliary dataset. After fixing its network parameters, the reconstructed EEG power spectral density map is input into the substitute classification network, and the semantic constraint loss between the network output and the real target state label is calculated. The specific steps are as follows:

[0064] An alternative classification network is pre-trained using the auxiliary dataset. The network parameters of the alternative classification network are iteratively optimized by minimizing the classification error, using the real EEG power spectral density map in the auxiliary dataset as input and the real target state label as reference.

[0065] With the network parameters of the alternative classification network fixed, the reconstructed EEG power spectral density map is input into the alternative classification network, and the semantic constraint loss between the network output and the real target state label is calculated. The specific calculation formula is as follows:

[0066] ,

[0067] in, Represents semantic constraint loss. This represents the EEG power spectral density map generated by the reconstruction. Indicates the actual target state label. This represents the total number of categories of the target state. Indicates category index, The label representing the true target state is a category. One-hot encoding value at time, This indicates that the reconstructed EEG power spectral density map belongs to the category. The semantic probability.

[0068] Step 6 (corresponding to S6): Combine the reconstruction loss function with the semantic constraint loss to construct the target total loss function; update the network parameters of the frequency domain feature generator based on the target total loss function to reconstruct and generate the final EEG power spectral density map. The specific steps are as follows:

[0069] Keeping the network parameters of the alternative classification network fixed, the partial derivative of the semantic constraint loss with respect to the reconstructed EEG power spectral density map is calculated, and the partial derivative is backpropagated to the frequency domain feature generator as a feedback gradient based on the chain rule.

[0070] The reconstruction loss function and the semantic constraint loss are weighted and jointly constructed to form a target total loss function. On the auxiliary dataset, the network parameters of the frequency domain feature generator are iteratively updated by minimizing this target total loss function. The final EEG power spectral density map is then reconstructed using the updated frequency domain feature generator. The specific formula for calculating the target total loss function is as follows:

[0071] ,

[0072] in, This represents minimizing the network parameters of the frequency domain feature generator. This represents the network parameters of the frequency domain feature generator. Represents the mathematical expectation. Represents the auxiliary dataset, Represents the reconstruction loss function. This represents the frequency domain feature generator function. This indicates the deep compression feature of the electroencephalogram (EEG). This represents the true power spectral density map of brainwaves. This represents the preset semantic constraint loss weight coefficients. Represents semantic constraint loss. This indicates the actual target status label.

[0073] Step 7 (corresponding to S7): Input the EEG power spectral density maps generated in steps S3 and S6, along with the actual EEG power spectral density maps, into the target classification network with fixed parameters, and calculate the semantic consistency rate of the target state. The specific steps are as follows:

[0074] The initial EEG power spectral density map generated in step S3 and the corresponding real EEG power spectral density map are respectively input into the target classification network with fixed parameters for prediction, and the proportion of samples that output the same target state prediction result is counted and used as the basic scene semantic consistency rate.

[0075] The final EEG power spectral density map generated in step S6 and the corresponding real EEG power spectral density map are respectively input into the target classification network with fixed parameters for prediction, and the proportion of samples that output the same target state prediction result is counted and used as the cross-domain scene semantic consistency rate.

[0076] like Figure 4 As shown, Figure 4 This is a visualization comparison of clustered scatter plots of EEG features in the network feature space. The figure shows that the reconstructed features in the target classification network promote the formation of clear decision boundaries between the baseline group and the target group, effectively improving the separability of the feature space and demonstrating the successful preservation and reconstruction of core semantic information.

[0077] like Figure 5 As shown, Figure 5 The bar chart shows the semantic consistency rate (SAR) of target state in the basic scenario and the cross-domain scenario. By introducing cross-domain semantic constraints, the semantic consistency rate in the cross-domain scenario has been significantly improved and maintained at a high level.

[0078] Step 8 (corresponding to S8): Combining the consistency rate, the identity tracing evaluation index, and the map reconstruction fidelity index, output the comprehensive evaluation result of the feature information of the deep EEG compression feature. The specific steps are as follows:

[0079] Extract the quantitative values ​​of the identity tracing evaluation index, the quantitative values ​​of the graph reconstruction fidelity index, and calculate the basic scene semantic consistency rate and the cross-domain scene semantic consistency rate.

[0080] Preset comprehensive evaluation weight coefficients are assigned to identity tracing, graph reconstruction fidelity, and semantic consistency, respectively. The semantic consistency rates of the basic scenario and the cross-domain scenario are then substituted into the linear weighted risk assessment model to calculate the comprehensive information evaluation scores for the basic scenario and the cross-domain scenario. The specific calculation formulas are as follows:

[0081] ,

[0082] ,

[0083] in, This indicates the comprehensive information assessment score for the basic scenario. This indicates the comprehensive information evaluation score for cross-domain scenarios. , and These represent the preset comprehensive evaluation weight coefficients corresponding to the identity tracing evaluation index, the graph reconstruction fidelity index, and the semantic consistency rate, respectively. This represents the quantitative value of the identity tracing and evaluation index. This represents the quantitative value of the fidelity index for the reconstructed map. This represents the semantic consistency rate of the basic scenario. This represents the semantic consistency rate of the cross-domain scenario. In a preferred embodiment of the present invention, based on the progressive harm caused by the overexposure of deep feature information by identity, physical form, and target state semantics, the preset risk assessment weight coefficient is preferably set as follows: , , .

[0084] The comprehensive information evaluation scores for the basic scenario and the cross-domain scenario are compared with a preset information retention threshold, respectively, and the comprehensive information evaluation results for the deep EEG compression features in the basic scenario and the cross-domain scenario are output. In a preferred embodiment of the present invention, the preset information retention threshold is preferably set to 0.85. When the calculated comprehensive information evaluation score is greater than or equal to the preset information retention threshold, the system determines that the feature compression has failed to achieve secure desensitization and outputs a comprehensive evaluation result of "significantly high retention of original sensitive data"; conversely, if the score is lower than the threshold, the system determines that the edge network constitutes an effective information bottleneck and outputs a comprehensive evaluation result of "effective data desensitization protection capability".

[0085] The above are merely one example of the implementation of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating deep EEG feature information based on frequency domain reconstruction, characterized in that, Includes the following steps: Step S1: Obtain the deep compression features of EEG transmitted in the communication channel and the corresponding real EEG power spectral density map. The deep compression features of EEG are generated by the end device inputting the original EEG signal into the feature coding network to extract the spatiotemporal high-dimensional features, and then compressing the spatiotemporal high-dimensional features into a low-dimensional vector space through the dimension reduction mapping layer. Step S2: Input the deep compression features of EEG into the pre-trained identity inference network, calculate the probability of the deep compression features of EEG belonging to the identity of each candidate subject, and determine the identity prediction accuracy based on the probability ranking, and use it as the identity tracing evaluation index. Step S3: Input the deep compression features of the EEG into the frequency domain feature generator to reconstruct and generate the corresponding EEG power spectral density map; Step S4: Based on the reconstructed EEG power spectral density map and the real EEG power spectral density map, calculate the mean square error and cosine similarity as the map reconstruction fidelity index, and construct a reconstruction loss function based on the mean square error and the cosine similarity. Step S5: In cross-domain scenarios where there are differences in data features between the source data and the auxiliary dataset, the alternative classification network is pre-trained using the auxiliary dataset. After fixing its network parameters, the reconstructed EEG power spectral density map is input into the alternative classification network, and the semantic constraint loss between the network output and the real target state label is calculated. Step S6: Combine the reconstruction loss function with the semantic constraint loss to construct the target total loss function; The network parameters of the frequency domain feature generator are updated based on the target total loss function to reconstruct and generate the final EEG power spectral density map. Step S7: Input the EEG power spectral density maps generated in Step S3 and Step S6, along with the actual EEG power spectral density maps, into the target classification network with fixed parameters, and calculate the semantic consistency rate of the target state. Step S8: Based on the consistency rate, the identity tracing evaluation index, and the map reconstruction fidelity index, output the comprehensive evaluation result of the feature information of the deep compression features of the EEG.

2. The method according to claim 1, characterized in that, Step S2 includes the following steps: Step 1: Based on the raw EEG signal features with subject labels in the auxiliary dataset, the identity inference network is pre-trained; Step 2: Input the deep EEG compression features into the pre-trained identity inference network, and calculate the probability that the deep EEG compression features belong to the identities of each candidate subject. The specific calculation formula is as follows: ， in, Indicates belonging to the first The probability of each candidate's identity. This indicates the deep compression feature of the electroencephalogram (EEG). The output of the identity inference network represents the first... The log odds of each candidate subject Indicates the first The log odds of a candidate subject; Step 3: Sort the identity probabilities in descending order of their numerical values, and select the top-ranked probabilities. Candidate subjects are selected to form a prediction set; if the real subject corresponding to the deep compression feature of the EEG exists in the prediction set, it is determined as a single prediction hit, and the Top-K identity prediction accuracy is calculated based on the number of hits, which is used as an evaluation indicator for identity tracing. It is a positive integer.

3. The method according to claim 1, characterized in that, Step S3 includes the following steps: Using the aforementioned deep compression features of EEG as input, a frequency domain feature generator network containing non-negative activation functions is constructed. The generator performs dimensionality analysis and expansion on the input compression features to generate a non-negative numerical matrix that matches the total number of electrode channels and the total number of frequency bands in the real EEG signal, which serves as the initial reconstructed EEG power spectral density map.

4. The method according to claim 1, characterized in that, Step S5 includes the following steps: Step 1: Pre-train an alternative classification network using the auxiliary dataset. Take the real EEG power spectral density map in the auxiliary dataset as input and the real target state label as reference. Iteratively optimize the network parameters of the alternative classification network by minimizing the classification error. Step 2: Fix the network parameters of the alternative classification network, input the reconstructed EEG power spectral density map into the alternative classification network, and calculate the semantic constraint loss between the network output and the real target state label. The specific calculation formula is as follows: ， in, Represents semantic constraint loss. This represents the EEG power spectral density map generated by the reconstruction. Indicates the actual target state label. This represents the total number of categories of the target state. Indicates category index, The label representing the true target state is a category. One-hot encoding value at time, This indicates that the reconstructed EEG power spectral density map belongs to the category. The semantic probability.

5. The method according to claim 1, characterized in that, Step S6 includes the following steps: Step 1: Keep the network parameters of the alternative classification network fixed, calculate the partial derivative of the semantic constraint loss with respect to the reconstructed EEG power spectral density map, and backpropagate the partial derivative as a feedback gradient to the frequency domain feature generator based on the chain rule. Step 2: The reconstruction loss function and the semantic constraint loss are weighted and jointly constructed to form a target total loss function. On the auxiliary dataset, the network parameters of the frequency domain feature generator are iteratively updated by minimizing this target total loss function. The final EEG power spectral density map is then reconstructed using the updated frequency domain feature generator. The specific formula for calculating the target total loss function is as follows: ， in, This represents minimizing the network parameters of the frequency domain feature generator. This represents the network parameters of the frequency domain feature generator. Represents the mathematical expectation. Represents the auxiliary dataset, Represents the reconstruction loss function. This represents the frequency domain feature generator function. This indicates the deep compression feature of the electroencephalogram (EEG). This represents the true power spectral density map of brainwaves. This represents the preset semantic constraint loss weight coefficients. Represents semantic constraint loss. This indicates the actual target status label.

6. The method according to claim 1, characterized in that, Step S7 includes the following steps: Step 1: Input the initial EEG power spectral density map generated in step S3 and the corresponding real EEG power spectral density map into the target classification network with fixed parameters for prediction, and count the proportion of samples that output the same target state prediction results, and use it as the basic scene semantic consistency rate. Step 2: Input the final EEG power spectral density map generated in step S6 and the corresponding real EEG power spectral density map into the target classification network with fixed parameters for prediction, and count the proportion of samples that output the same target state prediction results, and use it as the cross-domain scene semantic consistency rate.

7. The method according to claim 1, characterized in that, Step S8 includes the following steps: Step 1: Extract the quantitative values ​​of the identity tracing evaluation index, the quantitative values ​​of the graph reconstruction fidelity index, and the calculated basic scene semantic consistency rate and cross-domain scene semantic consistency rate; Step 2: Assign preset risk assessment weight coefficients to identity tracing, graph reconstruction fidelity, and semantic consistency; substitute the basic scenario semantic consistency rate and the cross-domain scenario semantic consistency rate into the linear weighted risk assessment model to calculate the comprehensive information assessment score for the basic scenario and the comprehensive information assessment score for the cross-domain scenario. The specific calculation formula is as follows: ， ， in, This indicates the comprehensive information assessment score for the basic scenario. This indicates the comprehensive information assessment score for cross-domain scenarios. , and These represent the preset risk assessment weight coefficients corresponding to the identity tracing assessment index, the graph reconstruction fidelity index, and the semantic consistency rate, respectively. This represents the quantitative value of the identity tracing and evaluation index. This represents the quantitative value of the fidelity index for the reconstructed map. This represents the semantic consistency rate of the basic scenario. This represents the semantic consistency rate of the cross-domain scenario; Step 3: Compare the comprehensive information evaluation score of the basic scenario and the comprehensive information evaluation score of the cross-domain scenario with the preset information retention threshold, and output the comprehensive information evaluation results of the deep compression features of EEG in the basic scenario and the cross-domain scenario, respectively.

Images