AIGC deepfake detection method and system based on dual-brain neural synchronous coupling

Abstract

The present invention discloses an AIGC deepfake detection method based on dual-brain neural synchronous coupling, comprising the following steps: S1: designing a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collecting dual-person EEG signals, and constructing a dual-person EEG database under the stimulation of real and AIGC-generated multimedia materials; S2: constructing a deep separable neural network based on the attention mechanism; S3: filtering the collected dual-person EEG signals to obtain signals in four different frequency bands, calculating the phase lock value (PLV), stacking them to obtain a three-dimensional vector, and inputting it into the deep separable neural network based on the attention mechanism to predict the authenticity of the content of the multimedia stimulus material. Also disclosed is an AIGC deepfake detection system based on dual-brain neural synchronous coupling. The present invention fully utilizes the information gain of dual-person EEG and the neural synchronization algorithm to analyze the dual-brain neural synchronous response, realizing AIGC deepfake detection inspired by the neural coupling mechanism.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence generated content (AIGC) authenticity detection technology at the intersection of computer ethics, brain-computer interface, and information security, and in particular to an AIGC deepfake detection method and system based on dual-brain neural synchronous coupling. Background Art

[0002] Currently, AIGC (Artificial Intelligence Generated Content) authenticity testing relies primarily on the following technical means, but these methods have significant limitations:

[0003] 1. Traditional visual / auditory detection methods: These rely on machine learning models to analyze the pixel or spectral features of images / audio, such as GAN generation trace detection. Traditional methods typically rely on hand-crafted features or simple machine learning models to extract image or audio features. These features often fail to capture sufficient details and patterns when faced with complex AIGC-generated content, resulting in reduced detection accuracy. With the rapid development of AIGC technology, new generative models such as diffusion models are constantly emerging, and traditional methods are struggling to cope with these new technologies.

[0004] 2. Single-subject (single-person) EEG detection: The neural response differences to AIGC are detected by decoding individual EEG signals, such as classification based on event-related potentials or spectral features. Although studies have used single-subject EEG for authenticity detection, this method has certain limitations. Single-subject detection cannot capture the interactive effects and group dynamics between multiple people. In a social environment with frequent interpersonal communication, the authenticity detection of AIGC needs to consider the interaction and emotional communication between multiple people, and single-subject EEG detection is difficult to provide this information. Faced with AIGC-generated materials, the human eye cannot distinguish between true and false, but there may be differences in neuronal responses and induced brain waves. The EEG of two people plays the role of data fusion and information gain, and the coupling of two people's neural synchronous responses becomes the key distinguishing information.

[0005] Therefore, there is an urgent need to provide a new AIGC deepfake detection method and system based on dual-brain neural synchronous coupling to solve the above problems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an AIGC deep fake detection method and system based on dual-brain neural synchronous coupling, which can enhance the frequency band representation of EEG signals, capture the interactive response and dynamics of two-person neural networks, and provide a technology for detecting the authenticity of multimedia information with the help of dual-brain response and deep learning.

[0007] To solve the above technical problems, the present invention adopts a technical solution: to provide an AIGC deepfake detection method based on dual-brain neural synchronous coupling, comprising the following steps:

[0008] S1: Design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect EEG signals of two people, and build a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials;

[0009] S2: Construct a deep separable neural network based on the attention mechanism, combining the attention mechanism with deep separable convolution to process EEG signal features;

[0010] S3: The collected EEG signals of the two people are filtered to obtain signals of four different frequency bands, and then the phase locking value PLV is calculated for the four frequency band signals. Finally, the four frequency band PLVs are spatially stacked to obtain a three-dimensional vector which is input into the deep separable neural network based on the attention mechanism to predict the authenticity of the hyperscan data content.

[0011] In a preferred embodiment of the present invention, in step S1, the design method of the two-person hyperscanning experimental paradigm includes:

[0012] Experimental conditions were manipulated using three factors: modality, valence, and life type. The modality factor included three levels: image, text, and audio. The valence condition included three levels: positive, neutral, and negative. The life type factor distinguished between living and non-living objects based on the nature of the stimulus. Living experimental objects included human faces, human voices, and text containing dialogues. Non-living stimulus materials included photos of everyday objects, natural sounds, and texts without any description of people or dialogues.

[0013] The main task of the participants was to evaluate the authenticity of various materials presented, including images, texts, and audio clips. Participants were asked to identify whether each stimulus was real or fabricated by AIGC based on their subjective evaluation.

[0014] Furthermore, the two-person hyperscanning experimental paradigm uses inter-subject correlation (ISC) to measure the synchronization of brain responses of two people when they receive the same external audiovisual stimulation, thereby verifying the effectiveness of the experimental paradigm design.

[0015] In a preferred embodiment of the present invention, in step S2, the method for constructing a deep separable neural network based on the attention mechanism includes:

[0016] S201: The input vector X1 is first globally average pooled It is a one-dimensional vector of size 1*1*C, where C is the feature dimension of X1. The average pooled one-dimensional vector is then put into two fully connected layers to learn the nonlinear relationship between channels. In the last layer, the weight vector of each channel is generated by the Sigmoid activation function. Finally, the weight vector and the original feature map are multiplied channel by channel to obtain a weighted N*N*4 feature vector, namely X2, where N is the number of electrodes.

[0017] S202: Convolve the four dimensions of the X2 feature vector with four 1*N convolution kernels respectively to obtain an N*1*4 feature vector X3;

[0018] S203: Convolve the feature vector X3 with four convolution kernels of size 1*1 to obtain the feature vector X4;

[0019] S204: Flatten the feature vector X4 to obtain a (N*4)*1 vector, input this vector into the softmax classifier for binary classification, and obtain 0 or 1, which indicates the authenticity of the EEG induced by the multimedia stimulation material.

[0020] In a preferred embodiment of the present invention, in step S3, the phase locking value PLV is selected as the synchronization indicator of the deep separable neural network model based on the attention mechanism, and the calculation method of the phase locking value PLV is:

[0021] First, calculate the phase locking value (PLV) of each frequency band at each time point for each subject:

[0022]

[0023] Where θ1(t, n) and θ2(t, n) are the instantaneous phases of the nth channel for each subject, N is the number of channels, and j is the imaginary unit;

[0024] The total PLV over time was then averaged across all time points.

[0025] To solve the above technical problems, another technical solution adopted by the present invention is to provide an AIGC deepfake detection system based on dual-brain neural synchronous coupling, comprising:

[0026] The two-person AIGC EEG database construction module is used to design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect two-person EEG signals, and construct a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials;

[0027] A deep separable neural network building module based on the attention mechanism is used to build a deep separable neural network based on the attention mechanism, combining the attention mechanism with deep separable convolution to process EEG signal features;

[0028] The data authenticity prediction module is used to filter the collected EEG signals of two people to obtain signals in four different frequency bands, then calculate the phase locking value PLV of the four frequency band signals, and finally spatially stack the four frequency band PLVs to obtain a three-dimensional vector which is input into the deep separable neural network based on the attention mechanism to predict the authenticity of the hyperscan data content.

[0029] In a preferred embodiment of the present invention, the method for designing a two-person hyperscanning experimental paradigm by constructing a two-person AIGC EEG database module includes:

[0030] Experimental conditions were manipulated using three factors: modality, valence, and life type. The modality factor included three levels: image, text, and audio. The valence condition included three levels: positive, neutral, and negative. The life type factor distinguished between living and non-living objects based on the nature of the stimulus. Living experimental objects included human faces, human voices, and text containing dialogues. Non-living stimulus materials included photos of everyday objects, natural sounds, and texts without any description of people or dialogues.

[0031] The participants' main task was to evaluate the authenticity of various materials presented, including images, text, and audio clips. Participants were asked to identify whether each stimulus was real or generated by AIGC based on their subjective evaluation.

[0032] Furthermore, the two-person hyperscanning experimental paradigm uses inter-subject correlation (ISC) to measure the synchronization of brain responses of two people when they receive the same external audiovisual stimulation, thereby verifying the effectiveness of the experimental paradigm design.

[0033] In a preferred embodiment of the present invention, the attention-based deep separable neural network constructed by the attention-based deep separable neural network building module includes:

[0034] Channel attention mechanism module is used to first globally average pool the input vector X1 It is a one-dimensional vector of size 1*1*C, where C is the feature dimension of X1. The average pooled one-dimensional vector is then put into two fully connected layers to learn the nonlinear relationship between channels. In the last layer, the weight vector of each channel is generated by the Sigmoid activation function. Finally, the weight vector and the original feature map are multiplied channel by channel to obtain a weighted N*N*4 feature vector, namely X2, where N is the number of electrodes.

[0035] A deep convolution module is used to convolve the four dimensions of the X2 feature vector output by the channel attention mechanism module with four convolution kernels of 1*N respectively to obtain an N*1*4 feature vector X3;

[0036] A point-by-point convolution module is used to convolve the feature vector X3 output by the depthwise convolution module with four convolution kernels of 1*1 size to obtain a feature vector X4;

[0037] A flattening module, configured to flatten the feature vector X4 output by the point-by-point convolution module to obtain a (N*4)*1 vector;

[0038] The classifier module is used to input the (N*4)*1 vector output by the flattening module into the softmax classifier for binary classification, and obtain 0 or 1, which is the authenticity of the EEG induced by the multimedia stimulation material.

[0039] In a preferred embodiment of the present invention, the method for the predicted data authenticity module to calculate the phase lock value PLV is:

[0040] First, calculate the phase locking value (PLV) of each frequency band at each time point for each subject:

[0041]

[0042] where θ1(t,n) and θ2(t,n) are the instantaneous phases of the nth channel for each subject, N is the number of channels, and j is the imaginary unit.

[0043] The total PLV over time was then averaged across all time points.

[0044] The beneficial effects of the present invention are:

[0045] (1) This paper proposes an AIGC deep fake content recognition technology based on a two-person EEG synchronized hyperscanning experimental paradigm, neural synchronization decoupling, and an improved deep separable neural network. By collecting EEG signals from two people, a EEG database of two people's synchronized recordings under the stimulation of real and AIGC fake materials is constructed, so that EEG detection can capture the interactive effect between two people and accurately detect AIGC fake materials.

[0046] (2) This paper develops an advanced neural architecture, namely a deep separable neural network based on the attention mechanism, which combines the attention mechanism with deep separable convolution to process PLV features. The adaptive reweighting amplifies the feature channels relevant to the task while attenuating those that are less important to the classification task. After the attention mechanism-driven reweighting, the refined feature map is fed into the deep separable convolution layer. This efficient convolution architecture separates spatial filtering and channel filtering, enhancing the network's ability to capture subtle inter-frequency dependencies while maintaining computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1This is a flow chart of the AIGC deepfake detection method based on dual-brain neural synchronous coupling of the present invention;

[0048] Figure 2 This is a diagram of the double-person hyperscanning experimental paradigm designed by the present invention;

[0049] Figure 3 This is the structural diagram of the deep separable neural network based on the attention mechanism;

[0050] Figure 4 This is a schematic diagram of the inter-subject correlation (ISC) results of the four frequency bands under different modalities between the real task and the AIGC task;

[0051] Figure 5 This is a schematic diagram of the inter-subject correlation (ISC) results under different task conditions based on valence;

[0052] Figure 6 It is a schematic diagram of the inter-subject correlation (ISC) results based on the classification of living and non-living materials;

[0053] Figure 7 This is a structural block diagram of the AIGC deepfake detection system based on dual-brain neural synchronous coupling. DETAILED DESCRIPTION

[0054] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0055] See also Figure 1 , embodiments of the present invention include:

[0056] An AIGC deepfake detection method based on dual-brain neural synchronous coupling includes the following steps:

[0057] S1: Design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect EEG signals of two people, and build a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials;

[0058] like Figure 2The experimental conditions shown were manipulated by three key factors: (1) modality, (2) valence, and (3) life type. The main task of the participants was to evaluate the authenticity of various materials presented to them. These materials included images, texts, and audio clips, which could be real or generated by AIGC. Participants were asked to identify whether each stimulus was real or fabricated by AIGC based on their subjective evaluation. The modality factor included three levels: image, text, and audio. The valence condition included three levels: positive, neutral, and negative. The life type factor distinguished between living and non-living objects based on the nature of the stimulus. Living experimental objects included human faces, human voices, and texts containing dialogues. Non-living stimulus materials included photos of everyday objects, natural sounds (such as environmental sounds), and texts without any description of people or dialogues. Therefore, the experiment adopted a 3 (modality: image, text, audio) × 3 (valence: positive, neutral, negative) × 2 (life type: living, non-living) design. The above-mentioned real images, texts, and audios were all from open source datasets, and the AI ​​images, texts, and audios were all generated by open source models.

[0059] The present invention uses intersubject correlation (ISC) to measure the synchronization of participants' brain responses when viewing the same stimulus. This is particularly suitable for naturalistic stimulation paradigms that conform to real-life scenarios because it allows measuring the consistency of neural responses between two individuals to natural stimulus multimedia materials without the need to predefine the stimulus events to appear periodically in the form of event locking, which is a typical event-related potential (ERP) paradigm for cognitive psychology research in laboratory settings. First, for each pair of participants k and l, the cross-electrode covariance matrix is ​​calculated. , which is:

[0060]

[0061] Among them, X k (t) and X l (t) represents the EEG signals of two individuals k and l at time point t. k and μ l Represents the mean of individuals k and l in the entire time series, usually the average of each individual signal at all time points. T is the total number of time points of the signal. For each time point t, calculate the signal X of each individual k (t) and X l (t) and their respective mean μ k and μ lThe product of these two deviations is summed to reflect the synchronization of the changes of the two individual signals at each time point. The covariance C is obtained by averaging the products of all time points. kl .

[0062] Then, the ISC value for each pair of participants was calculated using the following formula:

[0063]

[0064] here, is the i-th eigenvector, typically from principal component analysis (PCA) or other feature extraction methods, used to characterize the signal. T is the total number of time points, typically the length of the signal or the number of time frames in the experiment. For each pair of time points t, the dot product X of each individual signal and the eigenvector is calculated. k (t)*Vi and X l (t)*V i , respectively, sum the squares of these two dot products (numerator and denominator) to obtain their respective weighted signal intensities. The numerator is the sum of the weighted products of the signals of the two individuals at each time point, reflecting the synchronization of their signals. The denominator is the normalization of the product of the weighted square sum of the signals for standardization. Finally, the result of ISC is the ratio of the two, reflecting the given eigenvector V. i The degree to which two individuals are synchronized in time.

[0065] Through the study of inter-subject correlation (ISC), it was found that the inter-subject correlation (ISC) corresponding to the four frequency bands (δ, θ, α, β) under different conditions showed significant differences, which indicates that the four frequency bands show different neural synchronization responses when the brain distinguishes true and false materials at the level of social cognitive interpersonal neuroscience.

[0066] S2: Construct a deep separable neural network based on the attention mechanism, combining the attention mechanism with deep separable convolution to process EEG signal features;

[0067] The present invention selects PLV (Phase Locking Value) as the synchronization index of the model:

[0068]

[0069] where θ1(t, n) and θ2(t, n) are the instantaneous phases of the nth channel for each subject, N is the number of channels, and j is the imaginary unit. Finally, the total PLV of the experiment was averaged across all time points. PLV values ​​range from 0 to 1, with higher values ​​indicating greater synchronization.

[0070] S3: The collected EEG signals of the two people are filtered to obtain signals of four different frequency bands, and then the phase locking value PLV is calculated for the four frequency band signals. The PLV calculated for each frequency band is a matrix with a size of N*N. Finally, the four frequency band PLVs are spatially stacked to obtain a three-dimensional vector of N*N*4, which is input into the deep separable neural network based on the attention mechanism to predict the authenticity of the hyperscan data content. In this example, N is 32.

[0071] Specifically, the attention-based deep separable neural network (SEDSC) is as follows Figure 3 As shown in

[15] , it combines the attention mechanism with depthwise separable convolution to process these PLV features.

[0072] In the channel attention mechanism module, the input vector X1 is first globally average-pooled (Squeeze, i.e., Fsq(·)) into a one-dimensional vector of size 1*1*C. This operation can aggregate spatial information and capture the global distribution characteristics of the channel, so that subsequent operations can adjust the channel weights based on the global receptive field (C is dimension 4); then the Fex(·, W) operation is performed, and the one-dimensional vector after average pooling is placed into two fully connected layers to learn the nonlinear relationship between channels. In the last layer, the weight vector of each channel is generated by the Sigmoid activation function, and W is the parameter matrix of the fully connected layer; at this point, a one-dimensional weight vector will be obtained, and finally it will be multiplied with the original feature map channel by channel to obtain a weighted N*N*4 feature vector, i.e., X2.

[0073] In the depthwise convolution module, the four dimensions of the X2 feature vector are convolved with four 1*N convolution kernels, resulting in an N*1*4 feature vector X3. X3 is then convolved with four 1*1 convolution kernels to obtain X4. X3 is then flattened to obtain a (N*4)*1 vector. This vector is then input into a softmax classifier for binary classification, resulting in a 0 or 1 (corresponding to false or true).

[0074] In this neural network, adaptive reweighting amplifies task-relevant feature channels while attenuating those less important for the classification task. After attention-driven reweighting, the refined feature maps are fed into depthwise separable convolutional layers. This efficient convolutional architecture separates spatial filtering from channel filtering, enhancing the network's ability to capture subtle inter-frequency dependencies while maintaining computational efficiency.

[0075] The three conditions (modality, valence, and life type) designed using the method described in this invention all showed significant test results:

[0076] (1) Inter-subject correlation of four frequency bands under different modalities

[0077] Figure 4 Significant differences in inter-subject correlations (ISCs) between the real and AIGC tasks were observed, with distinct activation patterns observed across conditions. In both the text and audio tasks, ISCs were consistently higher in the real tasks than in the AIGC tasks. Specifically, ISC values ​​were significantly greater in the beta and alpha bands for the real tasks than in the AIGC tasks (p < 0.0001). In the audio task, ISCs differed in both low- and high-frequency bands. The difference was significant in the delta band (p < 0.01) and the beta band (p < 0.05). However, in the image task, the pattern was reversed: ISCs were significantly higher during the AIGC task than in the real task, particularly in the beta band (p < 0.01).

[0078] (2) Inter-subject correlations of the four frequency bands at different valences

[0079] Figure 5 The inter-subject correlations (ISCs) across different task conditions based on valence are shown. In the negative task, the ISCs induced by the realistic materials showed significant differences across all frequency bands, with the realistic materials consistently producing higher ISCs than the AIGC materials. The differences in ISCs in the beta band were highly significant (p < 0.01), and the differences in ISCs in the alpha, theta, and delta bands were all significant (p < 0.05). In the positive task, the ISCs induced by the realistic materials were significantly stronger in both the theta and alpha bands (p < 0.01), while those induced by the AIGC materials were weaker. In the neutral task, the ISCs induced by the realistic and AIGC materials were not significantly different (p > 0.05). However, across all tasks, the ISCs induced by the realistic materials were consistently higher than those induced by the AIGC materials.

[0080] (3) Inter-subject correlations of the four frequency bands at different valences

[0081] Figure 6 The inter-subject correlation (ISC) results for the classification of living and non-living materials are shown. The ISCs induced by real materials in the living group showed significant differences in both the theta and alpha bands, with significance levels of p < 0.05.

[0082] In addition, the network SEDSC (Queeze-and-Excitation-Enhanced Depth-

[0083] The Row-Filtered Depthwise Separable Convolutional Network (RDS) was compared with RFDSC (Row-Filtered Depthwise Separable Convolution Network) and the traditional CNN (convolution kernel size is 3) using five-fold cross validation, as shown in Table 1.

[0084] Table 1 Classification results of SEDSC compared with other models

[0085]

[0086] The proposed model achieved high classification accuracy across five different data input formats, consistently outperforming both CNN and RFDSC, with the highest accuracy reaching 92.42%. This improvement was nearly 10% across all frequency bands. This suggests that the inclusion of the attention mechanism enhances the representation of frequency bands, thereby improving the performance of SEDSC in classification tasks.

[0087] See Figure 7 The present invention also provides an AIGC deepfake detection system based on dual-brain neural synchronous coupling, including:

[0088] The two-person AIGC EEG database construction module is used to design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect two-person EEG signals, and construct a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials;

[0089] A deep separable neural network building module based on the attention mechanism is used to build a deep separable neural network based on the attention mechanism, combining the attention mechanism with deep separable convolution to process EEG signal features;

[0090] The data authenticity prediction module is used to filter the two-person EEG signals collected by the two-person AIGC EEG database construction module to obtain signals in four different frequency bands, then calculate the phase locking value PLV of the four frequency band signals, and finally stack the four frequency band PLVs spatially to obtain an N*N*4 three-dimensional vector, which is input into the deep separable neural network based on the attention mechanism to predict the authenticity of the content of multimedia stimulus materials.

[0091] The AIGC deep fake detection system based on dual-brain neural synchronous coupling in this example can execute the AIGC deep fake detection method based on dual-brain neural synchronous coupling provided by the present invention, can execute any combination of implementation steps of the method example, and has the corresponding functions and beneficial effects of the method.

[0092] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A deepfake detection method based on AIGC with dual-brain neural synchronous coupling, characterized in that: The following steps are involved: S1: Design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect EEG signals of two people, and build a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials; S2: Construct a deep separable neural network based on the attention mechanism, combining the attention mechanism with the deep separable convolution to process EEG signal features; the method for constructing the deep separable neural network based on the attention mechanism includes: S201: The input vector X1 is first globally average pooled It is a one-dimensional vector of size 1*1*C, where C is the feature dimension of X1. The average pooled one-dimensional vector is then put into two fully connected layers to learn the nonlinear relationship between channels. In the last layer, the weight vector of each channel is generated by the Sigmoid activation function. Finally, the weight vector and the original feature map are multiplied channel by channel to obtain a weighted N*N*4 feature vector, namely X2, where N is the number of electrodes. S202: Convolve the four dimensions of the X2 feature vector with four 1*N convolution kernels respectively to obtain an N*1*4 feature vector X3; S203: Convolve the feature vector X3 with four convolution kernels of size 1*1 to obtain the feature vector X4; S204: Flatten the feature vector X4 to obtain a (N*4)*1 vector, input this vector into the softmax classifier for binary classification, and obtain 0 or 1, indicating the authenticity of the EEG induced by the multimedia stimulus material; S3: Filter the EEG signals of the two people collected in step S1 to obtain signals of four different frequency bands, then calculate the phase locking value PLV of the four frequency band signals, and finally spatially stack the four frequency band PLVs to obtain a three-dimensional vector which is input into the deep separable neural network based on the attention mechanism constructed in step S2 to predict the authenticity of the content of the multimedia stimulus material.

2. The AIGC deepfake detection method based on dual-brain neural synchronous coupling according to claim 1 is characterized in that: In step S1, the design method of the two-person hyperscanning experimental paradigm includes: Experimental conditions were manipulated using three factors: modality, valence, and life type. The modality factor included three levels: image, text, and audio. The valence condition included three levels: positive, neutral, and negative. The life type factor distinguished between living and non-living objects based on the nature of the stimulus. Living experimental objects included human faces, human voices, and text containing dialogues. Non-living stimulus materials included photos of everyday objects, natural sounds, and texts without any description of people or dialogues. The main task of the participants was to evaluate the authenticity of various materials presented, including images, texts, and audio clips. Participants were asked to identify whether each stimulus was real or fabricated by AIGC based on their subjective evaluation.

3. The AIGC deepfake detection method based on dual-brain neural synchronous coupling according to claim 2 is characterized in that: The two-person hyperscanning experimental paradigm uses inter-subject correlation (ISC) to measure the synchronization of brain responses of two people when they receive the same external audiovisual stimulation, thereby verifying the effectiveness of the experimental paradigm design.

4. The AIGC deepfake detection method based on dual-brain neural synchronous coupling according to claim 1 is characterized in that: In step S3, the phase locking value PLV is selected as the synchronization index of the deep separable neural network model based on the attention mechanism. The calculation method of the phase locking value PLV is: First, calculate the phase locking value (PLV) of each frequency band at each time point for each subject: ， Where θ1(t, n) and θ2(t, n) are the instantaneous phases of the nth channel for each subject, N is the number of channels, and j is the imaginary unit; The total PLV over time was then averaged across all time points.

5. An AIGC deepfake detection system based on dual-brain neural synchronous coupling, characterized by: include: The two-person AIGC EEG database construction module is used to design a two-person hyperscanning experimental paradigm under the stimulation of digital multimedia materials, synchronously collect two-person EEG signals, and construct a two-person EEG database under the stimulation of real and AIGC-generated multimedia materials; A deep separable neural network building module based on an attention mechanism is used to build a deep separable neural network based on an attention mechanism, combining the attention mechanism with deep separable convolution to process EEG signal features; the deep separable neural network based on the attention mechanism includes: Channel attention mechanism module is used to first globally average pool the input vector X1 It is a one-dimensional vector of size 1*1*C, where C is the feature dimension of X1. The average pooled one-dimensional vector is then put into two fully connected layers to learn the nonlinear relationship between channels. In the last layer, the weight vector of each channel is generated by the Sigmoid activation function. Finally, the weight vector and the original feature map are multiplied channel by channel to obtain a weighted N*N*4 feature vector, namely X2, where N is the number of electrodes. A deep convolution module is used to convolve the four dimensions of the X2 feature vector output by the channel attention mechanism module with four convolution kernels of 1*N respectively to obtain an N*1*4 feature vector X3; A point-by-point convolution module is used to convolve the feature vector X3 output by the depthwise convolution module with four convolution kernels of 1*1 size to obtain a feature vector X4; A flattening module, configured to flatten the feature vector X4 output by the point-by-point convolution module to obtain a (N*4)*1 vector; A classifier module is used to input the (N*4)*1 vector output by the flattening module into a softmax classifier for binary classification, and obtain 0 or 1, indicating the authenticity of the EEG induced by the multimedia stimulation material; The data authenticity prediction module is used to filter the two-person EEG signals collected by the two-person AIGC EEG database construction module to obtain signals in four different frequency bands, then calculate the phase locking value PLV of the four frequency band signals, and finally spatially stack the four frequency band PLVs to obtain a three-dimensional vector which is input into the deep separable neural network based on the attention mechanism to predict the authenticity of the content of multimedia stimulus materials.

6. The AIGC deepfake detection system based on dual-brain neural synchronous coupling according to claim 5 is characterized in that: The method for designing a two-person hyperscanning experimental paradigm using the two-person AIGC EEG database construction module includes: Experimental conditions were manipulated using three factors: modality, valence, and life type. The modality factor included three levels: image, text, and audio. The valence condition included three levels: positive, neutral, and negative. The life type factor distinguished between living and non-living objects based on the nature of the stimulus. Living experimental objects included human faces, human voices, and text containing dialogues. Non-living stimulus materials included photos of everyday objects, natural sounds, and texts without any description of people or dialogues. The participants' main task was to evaluate the authenticity of various materials presented, including images, texts, and audio clips. Participants were asked to identify whether each stimulus was recorded from the objective real world or fabricated by AIGC based on their subjective evaluation.

7. The AIGC deepfake detection system based on dual-brain neural synchronous coupling according to claim 6 is characterized in that: The two-person hyperscanning experimental paradigm uses inter-subject correlation (ISC) to measure the synchronization of participants' brain responses when viewing the same stimulus, verifying the effectiveness of the experimental paradigm design.

8. The AIGC deepfake detection system based on dual-brain neural synchronous coupling according to claim 5 is characterized in that: The method for the prediction data authenticity module to calculate the phase lock value PLV is: First, calculate the phase locking value (PLV) of each frequency band at each time point for each subject: ， Where θ1(t, n) and θ2(t, n) are the instantaneous phases of the nth channel for each subject, N is the number of channels, and j is the imaginary unit; The total PLV over time was then averaged across all time points.

Images