Electroencephalogram fuzzy commitment authentication method and device thereof
By processing EEG signals through a densely connected neural network with a neural LDPC decoder and a channel-axis attention mechanism, the problems of signal non-stationarity and privacy security in EEG signal recognition are solved, and robust feature extraction and efficient authentication processes are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA NORMAL UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196974A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biometric identification and information security technology, and in particular to a method and apparatus for fuzzy commitment authentication using electroencephalogram (EEG) signals. Background Technology
[0002] With the development of brain science and neural engineering, identity recognition technology based on electroencephalogram (EEG) signals (i.e., "brainprint recognition") has shown great potential in high-security identity authentication scenarios due to its unforgeability, liveness detection characteristics, and unavailability under coercion.
[0003] However, despite the promising future of brainprint recognition technology, it faces two major challenges in practical applications: First, the non-stationarity and significant intra-class variability of the signal. EEG signals are weak and easily affected by psychological and physiological states such as emotions, fatigue, and attention. Signals collected from the same individual at different times exhibit significant differences (i.e., intra-class variation), leading to unstable extracted features. Existing feature extraction methods (such as traditional time-frequency analysis) struggle to extract robust "fingerprints" that remain robust under these variations. While methods based on wavelet scattering transform and channel-axis attention mechanisms have improved feature robustness to some extent, converting continuous deep features into stable discrete keys remains a challenge. Second, the privacy and security of biometric templates. Traditional biometric systems typically store users' original biometric templates in databases. If this database is leaked, the irreversible nature of biometrics risks permanently invalidating the user's EEG biometrics.
[0004] To address these issues, the Fuzzy Commitment Scheme (FCS) was proposed, which binds biometric features to a random key using error-correcting codes. However, existing error-correcting codes (such as BCH codes, RS codes, or standard LDPC decoding algorithms) typically assume that the noise is Gaussian white noise or burst noise. In contrast, the biometric differences in EEG signals exhibit highly nonlinear and complex distribution characteristics. Consequently, existing decoding algorithms cannot effectively correct feature errors caused by changes in physiological state, resulting in an excessively high rejection rate for legitimate users and failing to meet practical authentication requirements. Summary of the Invention
[0005] Based on this, the purpose of the present invention is to provide a method for fuzzy commitment authentication of EEG signals, which has the advantages of strong error correction redundancy, good error correction performance, and low rejection rate.
[0006] A method for fuzzy commitment authentication using electroencephalogram (EEG) signals includes the following steps:
[0007] Based on the raw EEG signals, obtain the time-frequency feature tensor; Based on the time-frequency feature tensor, a continuous feature vector containing human identity information is obtained; Based on continuous feature vectors, binary brain pattern biometrics are obtained; Generate high-entropy true random keys; Based on the high-entropy true random key, obtain the registration error correction codeword; Obtain the registration key hash based on the high-entropy true random key and store the registration key hash; Based on the registered error correction codeword and binary brainprint biometric features, obtain the fuzzy commitment value and store the fuzzy commitment value; Acquire binary brainprint biometrics of the EEG signal to be tested; Based on the binary brainprint biometrics and fuzzy commitment values of the EEG signal to be tested, obtain the sequence to be decoded; Based on the sequence to be decoded, obtain the authentication error correction codewords through the neural LDPC decoder; Based on the authentication error correction codeword, obtain the authentication random key and the authentication key hash corresponding to the authentication random key; based on the authentication key hash and the registration key hash, realize the authentication of the EEG signal to be tested; The neural LDPC decoder comprises multiple iterative layers, each iterative layer following the Tanner graph of the LDPC code, with learnable parameters set in the check node layer and / or variable node layer.
[0008] Compared to existing technologies, this invention uses a neural LDPC decoder for decoding and adaptively fits and tracks the complex noise distribution of EEG biometrics through learnable parameters, thereby achieving strong error correction redundancy, effectively tolerating differences in biometrics, reducing the rejection rate, and ensuring error correction performance.
[0009] Furthermore, in each of the iterative layers, residual connections and layer normalization are employed in the variable node layer.
[0010] Furthermore, a noisy pre-training sequence was constructed within the range of 1.5 dB to 4.5 dB of EEG biometric differences to pre-train the neural LDPC decoder.
[0011] Furthermore, the continuous feature vector is obtained by using a densely connected neural network with a channel-axis attention mechanism as input, with the time-frequency feature tensor as input. The densely connected neural network with the channel-axis attention mechanism contains a multi-dimensional channel-axis attention module. Each channel-axis attention module includes a channel attention branch and an axial spatial attention branch. The axial spatial attention branch includes an electrode attention sub-branch and a temporal attention sub-branch.
[0012] Furthermore, wavelet scattering transform is performed on the original EEG signal to obtain the time-frequency feature tensor.
[0013] Furthermore, the continuous feature vector is subjected to adaptive differential binarization to obtain the binary brainprint biometrics, wherein the adaptive differential binarization adopts a misalignment comparison and cyclic filling mechanism.
[0014] Further, an XOR operation is performed on the registered error correction codeword and the binary brainprint biometric feature to obtain the fuzzy commitment value; and an XOR operation is performed on the binary brainprint biometric feature of the EEG signal to be tested and the fuzzy commitment value to obtain the sequence to be decoded.
[0015] Furthermore, the binary brainprint biometrics is rearranged using a preset permutation sequence, and then an XOR operation is performed between the registered error correction codeword and the rearranged binary brainprint biometrics to obtain a fuzzy commitment value. In addition, the binary brainprint biometrics of the EEG signal to be tested is rearranged in the same way using the preset permutation sequence, and then an XOR operation is performed between the rearranged binary brainprint biometrics of the EEG signal to be tested and the fuzzy commitment value to obtain the sequence to be decoded.
[0016] Furthermore, the original EEG information is sequentially downsampled, filtered, subjected to independent component analysis, and subjected to common average reference to remove noise and obtain a denoised EEG signal; the denoised EEG signal is then subjected to wavelet scattering transform to obtain the time-frequency feature tensor.
[0017] In addition, the present invention also provides a brainwave signal fuzzy commitment authentication device, which includes a preprocessing module, a feature extraction module, a binary biometric feature acquisition module, a key generation module, an error correction codeword generation module, a key hash acquisition module, a fuzzy commitment value acquisition module, an authentication module, and a storage module; The preprocessing module is used to denoise the acquired raw EEG signals to obtain denoised EEG signals. The feature extraction module includes a time-frequency feature tensor acquisition module and a depth feature acquisition module; wherein, the time-frequency feature tensor acquisition module is used to obtain a time-frequency feature tensor based on the denoised EEG signal, and the depth feature acquisition module is used to obtain a continuous feature vector containing human identity information based on the time-frequency feature tensor; The binary biometric acquisition module is used to obtain binary brain pattern biometrics based on the continuous feature vector; The key generation module is used to generate high-entropy true random keys; The error correction codeword generation module is used to encode the high-entropy true random key into a registered error correction codeword; The key hash acquisition module is used to perform hash calculation on the high-entropy true random key to obtain the registration key hash; The fuzzy commitment value acquisition module is used to generate fuzzy commitment values based on the registered error correction codewords and binary brainprint biometric features; The authentication module is used to acquire the binary brainprint biometric features of the EEG signal to be tested, acquire the sequence to be decoded based on the binary brainprint biometric features and fuzzy commitment value of the EEG signal to be tested, acquire the authentication error correction codeword through the neural LDPC decoder based on the sequence to be decoded, acquire the authentication random key and the authentication key hash corresponding to the authentication random key based on the authentication error correction codeword, and acquire the authentication key hash based on the authentication key hash and the registration key hash to achieve authentication of the EEG signal to be tested; wherein, the neural LDPC decoder contains multiple iterative layers, each iterative layer follows the Tanner graph of the LDPC code, and learnable parameters are set in the verification node layer and / or variable node layer; The storage module is used to store the registration key hash and the fuzzy commitment value.
[0018] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the fuzzy commitment authentication method for EEG signals of the present invention. Figure 2 This is another flowchart illustrating the fuzzy commitment authentication method for EEG signals of the present invention; Figure 3 The present invention relates to a device for ambiguity commitment authentication of electroencephalogram (EEG) signals. Figure 4 This is a schematic diagram of the densely connected neural network structure of the channel-axial attention mechanism of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention more readily understood by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] refer to Figure 1-3 , Figure 1 A flowchart illustrating the fuzzy commitment authentication method for electroencephalogram signals of the present invention is shown. Figure 2 Another flowchart of the EEG signal fuzzy commitment authentication method of the present invention is shown. Figure 3 This invention relates to a device for fuzzy commitment authentication of electroencephalogram (EEG) signals. Specifically, the EEG signal fuzzy commitment authentication method of this invention includes two stages: fuzzy commitment registration and authentication. S1. Obtain continuous feature vectors containing human identity information based on the original EEG signals.
[0022] In this invention, during the feature extraction stage of EEG signals, feature extraction module 2 obtains a time-frequency feature tensor through time-frequency feature tensor acquisition module 2-1, and obtains a continuous feature vector containing human identity information based on the time-frequency feature tensor through deep feature acquisition module 2-2. Specifically, this includes: S11. Obtain the time-frequency feature tensor based on the original EEG signal.
[0023] In this step, the time-frequency feature tensor acquisition module 2-1 performs a wavelet scattering transform (WST) on the original EEG signal to obtain a time-frequency feature tensor with translation invariance. In this invention, the translation invariance and deformation stability of the scattering transform are used to resist the slight time shifts in the EEG signal during acquisition, thus effectively compensating for the shortcomings of traditional time-frequency analysis in processing transient features and providing a stable time-frequency input for deep feature extraction in subsequent steps.
[0024] In one specific embodiment, the scattering scale J=7 of the wavelet scattering transform and the number of wavelets per octave Q=8 are set. Through cascaded first- and second-order complex wavelet convolutions, modulo operations, and local averaging operations, the one-dimensional EEG signal is mapped into a scattering coefficient matrix, which is a time-frequency feature tensor with translation invariance. The setting of the number of wavelets per octave Q=8 in this embodiment ensures that there is sufficient filter density in the frequency dimension to capture subtle identity differences in key EEG rhythms while avoiding excessive computational overhead caused by feature dimension explosion.
[0025] S22. Based on the time-frequency feature tensor, obtain a continuous feature vector containing human identity information.
[0026] In this step, the deep feature acquisition module 2-2 uses a densely connected neural network (CAA-DenseNet) with a channel-axis attention mechanism to extract continuous feature vectors containing individual identity information, taking the time-frequency feature tensor as input.
[0027] Specifically, to effectively extract EEG signal features robust to spatiotemporal shifts, refer to Figure 4This invention employs a densely connected neural network with a channel-axial attention mechanism. It includes multi-dimensional channel-axial attention modules to filter key frequency bands and electrode information, thereby capturing deep identity associations of EEG signals in the electrode topology and time-frequency domain. Each channel-axial attention module includes a channel attention branch and an axial spatial attention branch. The axial spatial attention branch contains an electrode attention sub-branch and a temporal attention sub-branch. The channel attention branch aggregates global spatial information and performs weighted filtering on the channel dimension of the time-frequency feature tensor, highlighting key frequency band features. The axial attention models long-distance dependencies in both the time and electrode dimensions. Its electrode attention sub-branch extracts the spatial topological dependencies between different EEG electrodes by compressing the time dimension, and its temporal attention sub-branch extracts the dynamic temporal dependencies before and after a time point by compressing the electrode dimension.
[0028] In this invention, by combining channel weighting and axial spatial modeling in the attention configuration, the densely connected neural network can automatically filter stable features from EEG signals, even from high-noise EEG signals. Furthermore, the features selected through the attention mechanism exhibit strong cross-session robustness, reducing bit-flip errors in the binary brainprint biometric sequence obtained through binarization due to differences in physiological state during biometric acquisition, resulting in a more stable binary brainprint biometric sequence. Since the extracted features preserve the space-frequency topology of the EEG, the resulting bit-flip errors have a learnable pattern, allowing the neural LDPC decoder to more accurately train its learnable parameters, thereby improving the error correction success rate for complex biological variations. In addition, the channel-axial attention mechanism employed in this invention can filter out random fluctuations in EEG signals caused by factors such as emotion and fatigue, minimizing fluctuations in similar samples. This invention ensures that the characteristic variations of legitimate users under different physiological states always fall within the effective error correction range of the neural LDPC decoder, thereby reducing the rejection rate. At the same time, while compressing the random fluctuations of the same user, this mechanism also deeply explores the essential identity differences between different brain electrode topologies, "pushing far apart" the characteristics of different individuals in space. This ensures that the characteristic distance of illegitimate users far exceeds the error correction limit of the decoder, fundamentally eliminating erroneous acceptance and thus reducing the false recognition rate. Therefore, the method of this invention maximizes individual distinguishability while minimizing the fluctuations of similar samples, enabling high feature recognition rates to be achieved even in extremely noisy environments (such as 1.5dB), effectively balancing low false recognition and low rejection.
[0029] In one specific embodiment, a densely connected neural network based on the channel-axis attention mechanism of the DenseNet-121 architecture is adopted. A channel-axis attention module is embedded between two adjacent dense blocks to reweight the high-dimensional features in terms of channels and space, and a transition layer is used for pooling dimensionality reduction. This ensures that the features passed to the next dense block condense the key identity information while controlling the computational complexity. The growth rate k=32 of the DenseNet-121 architecture ensures efficient feature propagation. The output dimension is 260-dimensional.
[0030] In a preferred embodiment, the densely connected neural network of the channel-axis attention mechanism is pre-trained to optimize network parameters and improve the network's generalization performance. The specific pre-training method is not limited in this invention; those skilled in the art can set it according to actual needs.
[0031] In addition, refer to Figure 2 Considering that EEG signals are weak and easily interfered with, in order to improve the accuracy of brain pattern biometric extraction, the present invention preferably includes the following steps before feature extraction: Step S0: Denoise the raw EEG signal to obtain a denoised EEG signal.
[0032] Specifically, the preprocessing module 1 sequentially performs downsampling, filtering, independent component analysis (ICA), and common average reference on the original EEG signal to remove noise and obtain a denoised EEG signal. Then, in step S11, the time-frequency feature tensor acquisition module 2-1 performs wavelet scattering transform on the denoised EEG signal to obtain the time-frequency feature tensor.
[0033] In one specific embodiment, the preprocessing module 1 performs 128Hz downsampling and 0.5-42Hz bandpass filtering on the acquired raw EEG signal to remove power frequency interference and high-frequency electromyography noise. Then, in order to further improve the signal-to-noise ratio, the independent component analysis algorithm is used to automatically identify and remove electrooculography artifacts. After that, a common average reference is used to perform rereference processing on the multi-channel signal to eliminate common-mode noise, and finally a denoised EEG signal is obtained.
[0034] S2. Obtain binary brain pattern biometrics based on continuous feature vectors.
[0035] In this step, the binary biometric acquisition module 3 performs adaptive differential binarization on the continuous feature vector to obtain binary brainprint biometrics.
[0036] The adaptive differential binarization employs a misalignment comparison and cyclic padding mechanism to generate binary codes, forming a closed-loop feature chain. It utilizes the relative variation trend between feature points rather than absolute amplitudes to eliminate the influence of signal amplitude fluctuations caused by changes in electrode impedance, significantly improving the robustness of features across sessions and enhancing the cross-session stability of brain pattern signals. Specifically: For the first L 1 bit (i, j = 1, 2, ..., L) 1):
[0037] For the last element (i, j = L):
[0038] The continuous eigenvectors are represented as F = {f1, f2, ..., f...} L The generated binary brainprint biometrics is represented as B={b1, b2, ..., b}. L}, where f i b represents an element in a continuous eigenvector. j This represents an element in the binary biometric brain pattern, where L is the feature length.
[0039] S3. Generate a high-entropy true random key.
[0040] In this step, a high-entropy true random key is generated using key generation module 4 to ensure system security. The specific method for generating the high-entropy true random key is not limited in this invention; those skilled in the art can set it according to actual needs.
[0041] In this invention, the order of step S3 is not limited to after step S2; it only needs to be before the subsequent steps of obtaining the registration error correction codeword and the registration key hash.
[0042] S4. Obtain the registration error correction codeword based on the high-entropy true random key.
[0043] In this step, the error correction codeword generation module 5 uses a quasi-cyclic LDPC (Low-density Parity-check) encoder to encode the high-entropy true random key into error correction codewords, i.e., the registered error correction codewords. Given the high noise and high variability of EEG signals, low code rate encoding is preferred in this step. The generated low code rate error correction codewords can provide a very strong error correction redundancy space, providing a basis for effectively tolerating differences in biometrics.
[0044] This invention employs a quasi-cyclic LDPC encoder, which achieves high storage efficiency by storing only the encoder's base map and a single integer boost factor. Simultaneously, this encoder possesses powerful parallel computing capabilities, making it suitable for accelerating neural networks. Furthermore, this encoder can accurately match the output dimension of the extracted EEG features, enabling a one-to-one correspondence between error-correcting codewords and the corresponding dimensional data of the EEG features.
[0045] In one specific embodiment, a 50-bit high-entropy true random key is generated. The base graph of the quasi-cyclic LDPC encoder is BG2, which has 52 columns and 42 rows. The boost factor Z=5, and a 260-bit error correction codeword is generated.
[0046] S5. Obtain the registration key hash based on the high-entropy true random key and store the registration key hash.
[0047] In this step, the key hash acquisition module 6 obtains the registration key hash through hash calculation.
[0048] In this invention, the order of step S5 is not limited to after step S4; it can also be before step S4. S6. Based on the registered error correction codeword and binary brainprint biometrics, obtain the fuzzy commitment value and store the fuzzy commitment value.
[0049] In this step, the fuzzy commitment value acquisition module 7 performs an exclusive OR (XOR) operation on the aforementioned registration error correction code and binary brainprint biometric features to generate and store the fuzzy commitment value.
[0050] In a preferred embodiment, the binary brainprint biometric features are rearranged using a preset permutation sequence before the XOR operation. Then, the registered error correction codeword is XORed with the rearranged binary brainprint biometric features to generate a fuzzy commitment value, as follows: (1) Where δ represents the fuzzy commitment value, This indicates the error correction code. Represents binary brainprint biometrics. This represents the binary brainprint biometric features resulting from random rearrangement using a pre-defined permutation sequence P. This represents the XOR operation.
[0051] The scrambling in this step disperses potential bursts of continuous errors into random errors, enabling a more even distribution of these errors. This fully leverages the error correction threshold advantage of the neural LDPC decoder, significantly improving the success rate of key recovery.
[0052] In this invention, the XOR operation involves only bit-level logical calculations, resulting in extremely low computational complexity and high efficiency. This enables millisecond-level rapid response during the registration phase, significantly saving computational resources for the EEG acquisition terminal. Furthermore, without knowing the random key, the fuzzy commitment value generated by the XOR operation cannot be used to deduce the key or original brainprint features, thus protecting the user's biometric privacy. In the event of database leakage, the current fuzzy commitment value can be invalidated, mitigating the risk of permanent invalidation of biometric features once leaked. This provides high security, irreversibility, and revocability. Benefiting from the linear superposition property of the XOR operation, during the decryption process in the authentication phase, the feature differences (i.e., intra-class variations) collected from the same user at different times can be losslessly converted into bit-flip errors appended to the error correction codeword. This allows the true biometric noise to be completely transmitted to the neural LDPC decoder for adaptive error correction. The neural LDPC decoder can then use deep learning algorithms to accurately capture these difference patterns, achieving a much higher authentication pass rate than traditional hard-decision error correction (such as BCH codes). This also preserves the error distribution to adapt to neural error correction, further improving the authentication pass rate.
[0053] This completes the fuzzy commitment registration phase, specifically steps S0-S6. In this invention, after generating and storing the fuzzy commitment value and the registration key hash value, the original EEG information can be immediately destroyed. That is, the solution of this invention can achieve the goal of not storing the user's original EEG characteristics, but instead storing the fuzzy commitment value and the registration key hash, thus ensuring privacy and security.
[0054] S7. Authentication of the EEG signal to be tested is achieved based on the registration key hash and fuzzy commitment value.
[0055] When a user requests authentication, the authentication module 8 performs an authentication operation on the collected EEG signal to be tested, specifically including: S71. The EEG signal to be tested is processed through steps S1-S2 or steps S0-S2 to obtain the binary brainprint biometrics of the EEG signal to be tested.
[0056] S72. Perform an XOR operation between the binary brainprint biometric features of the EEG signal to be tested and the fuzzy commitment value obtained in step S6 above to obtain the sequence to be decoded. This sequence to be decoded includes authentication error correction codewords and biometric noise.
[0057] Furthermore, if the binary brainprint biometrics were rearranged using a preset permutation sequence in step S6, then step S72 is as follows: First, the same rearrangement operation is performed on the binary brainprint biometrics of the EEG signal to be tested using the preset permutation sequence in step S6 to randomly disperse sudden random errors; then, the rearranged binary brainprint biometrics of the EEG signal to be tested is XORed with the fuzzy commitment value obtained in step S6 to obtain the sequence to be decoded. Specifically: (2) in, D Indicates the sequence to be decoded. Binary brain striae biometrics representing the EEG signal to be measured. This represents the binary brainprint biometrics of the EEG signal to be tested obtained by rearranging the same preset permutation sequence P in step S6.
[0058] Combining formulas (1) and (2), it can be seen that the same pre-set permutation sequence is used in both the registration and authentication stages to rearrange the binary EEG biofeatures in the registration and authentication stages respectively. Thus, the sequence to be decoded can be represented in the form of authentication error correction codeword + random noise. Therefore, the neural LDPC decoder can directly use the check matrix for error correction during decoding without the need for "inverse permutation" related operations.
[0059] S73. Input the sequence to be decoded into the neural LDPC decoder to obtain the authentication error correction codeword.
[0060] In this step, the neural LDPC decoder contains multiple iterative layers. Each iterative layer follows the Tanner graph of the LDPC code. Learnable parameters are set in its variable node layer and / or check node layer to effectively correct the nonlinear bit flip error unique to EEG signals by adaptively learning the biometric noise distribution of EEG signals, tolerate greater biometric differences, and significantly reduce the resistance rate of legitimate users.
[0061] The neural LDPC decoder is preferably based on an expanded min-sum decoding network, containing 25 iterative layers for 25 iterations. This depth is sufficient to ensure error correction performance while avoiding excessive computational latency. Residual connections and layer normalization are introduced into the variable node layers within each iterative layer to suppress gradient vanishing and confidence explosion problems in deep iterations of noisy EEG features, resulting in higher stability and reducing the rejection rate. The verification node layers utilize the min-sum algorithm to process the verification logic, and the information passed to the variable nodes is: the result calculated by the min-sum algorithm multiplied by a weight parameter and then added to a bias parameter. In the hidden layer iterations, the aforementioned information passed to the variable nodes is used for the confidence update calculation of the next layer's variable nodes. In the output layer, the aforementioned information passed to the variable nodes is used for the final posterior probability calculation. The weight and bias parameters are learnable parameters, and the reliability weights and biases of the information passed to different verification nodes can be adjusted by adaptively learning the biometric noise distribution of the EEG signal.
[0062] Furthermore, if low bitrate encoding is adopted in step S4 of the registration phase, the neural LDPC decoder in this step can make full use of these redundant check bits to adaptively correct a high proportion of random bit flip errors, thereby achieving further effective tolerance for biometric differences and further improving the authentication accuracy.
[0063] S74. Based on the authentication error correction codeword, obtain the authentication random key and the authentication key hash of the authentication random key, compare the authentication key hash with the registration key hash, and complete the authentication.
[0064] In this step, authentication is successful if the authentication key hash matches the registration key hash; otherwise, it is rejected.
[0065] To improve the adaptability of the neural LDPC decoder to the non-stationary characteristics of EEG signals and thus enhance authentication accuracy, this invention pre-trains the neural LDPC decoder, forcing it to learn to adapt to different degrees of biometric differences, thereby adjusting the learnable parameters.
[0066] During pre-training, a Mixed Signal-to-Noise Ratio (Mixed-SNR) strategy is preferred for constructing the training dataset. This involves treating the intra-class variation of EEG biometrics as additive noise in the communication channel, and constructing noisy pre-training sequences within an EEG biometric difference range of 1.5 dB to 4.5 dB. The loss function is based on the Binary Cross-Entropy Loss (BCE Loss) with an added entropy regularization term. By minimizing the entropy of the predicted probability distribution, the decoder outputs a high-confidence (close to 0 or 1) binarized result, thereby improving the determinism of key recovery.
[0067] In addition, the present invention also provides a fuzzy commitment authentication device for electroencephalogram (EEG) signals, which includes a preprocessing module 1, a feature extraction module 2, a binary biometric feature acquisition module 3, a key generation module 4, an error correction codeword generation module 5, a key hash acquisition module 6, a fuzzy commitment value acquisition module 7, an authentication module 8, and a storage module 9.
[0068] The preprocessing module 1 is used to perform downsampling, filtering, independent component analysis, and common average reference sequentially on the acquired raw EEG signals to remove noise from the raw EEG signals and obtain denoised EEG signals.
[0069] The feature extraction module 2 includes a time-frequency feature tensor acquisition module 2-1 and a depth feature acquisition module 2-2; wherein, the time-frequency feature tensor acquisition module 2-1 is used to perform wavelet scattering transform on the original EEG signal or the denoised EEG signal to obtain a time-frequency feature tensor with translation invariance, and the depth feature acquisition module 2-2 is used to obtain a continuous feature vector containing human identity information by using a densely connected neural network with a channel-axis attention mechanism and the time-frequency feature tensor as input.
[0070] The binary biometric acquisition module 3 is used to perform adaptive differential binarization on continuous feature vectors to obtain binary brainprint biometrics.
[0071] The key generation module 4 is used to generate high-entropy true random keys.
[0072] The error correction codeword generation module 5 is used to encode the high-entropy true random key into a registration error correction codeword using a quasi-cyclic LDPC encoder.
[0073] The key hash acquisition module 6 is used to perform hash calculation on the high-entropy true random key to obtain the registration key hash.
[0074] The fuzzy commitment value acquisition module 7 is used to generate fuzzy commitment values based on the registered error correction codewords and binary brainprint biometric features.
[0075] The authentication module 8 is used to authenticate the EEG signal to be tested based on the registration key hash and fuzzy commitment value.
[0076] The storage module 9 is used to store the registration key hash and the fuzzy commitment value.
[0077] Compared to existing technologies, the neural LDPC decoder in this invention contains learnable parameters. By adaptively learning the biometric noise distribution of EEG signals, it adjusts the reliability parameters of information transmission between different verification nodes and / or variable nodes, effectively correcting the nonlinear bit-flipping errors unique to EEG signals, tolerating greater biometric differences, improving the success rate of error correction for complex biological variations, significantly reducing the resistance rate of legitimate users, and achieving high authentication accuracy. Furthermore, residual connections and layer normalization are introduced into the variable node layer of the neural LDPC decoder to suppress the gradient vanishing and confidence explosion problems of high-noise EEG features in deep iterations, resulting in high stability and further reducing the rejection rate. The channel-axis attention mechanism minimizes fluctuations in similar samples, ensuring that the feature variations of legitimate users under different physiological states always fall within the effective error correction range of the neural LDPC decoder, thereby reducing the rejection rate. Simultaneously, this mechanism also deeply mines the essential identity differences between different brain electrode topologies, ensuring that the features of illegitimate users are incorrectly accepted, reducing the false recognition rate.
[0078] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.
Claims
1. A method for fuzzy commitment authentication using electroencephalogram (EEG) signals, characterized in that, Includes the following steps: Based on the raw EEG signals, obtain the time-frequency feature tensor; Based on the time-frequency feature tensor, a continuous feature vector containing human identity information is obtained through a densely connected neural network with a channel-axis attention mechanism. Based on continuous feature vectors, binary brain pattern biometrics are obtained; Generate high-entropy true random keys; Based on the high-entropy true random key, the registered error correction codeword is obtained using a low-coding-rate quasi-cyclic LDPC encoder; Obtain the registration key hash based on the high-entropy true random key and store the registration key hash; Based on the registered error correction codeword and binary brainprint biometric features, obtain the fuzzy commitment value and store the fuzzy commitment value; Acquire binary brainprint biometrics of the EEG signal to be tested; Based on the binary brainprint biometrics and fuzzy commitment values of the EEG signal to be tested, obtain the sequence to be decoded; Based on the sequence to be decoded, obtain the authentication error correction codewords through the neural LDPC decoder; Based on the authentication error correction codeword, obtain the authentication random key and the authentication key hash corresponding to the authentication random key; The authentication of the EEG signal to be tested is achieved by using the authentication key hash and the registration key hash. The neural LDPC decoder comprises multiple iterative layers, each iterative layer following the Tanner graph of the LDPC code, with learnable parameters set in the check node layer and / or variable node layer.
2. The authentication method according to claim 1, characterized in that, In each of the iterative layers, residual connections and layer normalization are used in the variable node layer.
3. The authentication method according to claim 1 or 2, characterized in that, A noisy pre-training sequence was constructed within the range of 1.5 dB to 4.5 dB of EEG biometric differences to pre-train the neural LDPC decoder.
4. The authentication method according to claim 1, characterized in that, The densely connected neural network of the channel-axial attention mechanism contains a multi-dimensional channel-axial attention module. Each channel-axial attention module includes a channel attention branch and an axial spatial attention branch. The axial spatial attention branch includes an electrode attention sub-branch and a temporal attention sub-branch.
5. The authentication method according to claim 1, characterized in that, The time-frequency feature tensor is obtained by performing wavelet scattering transform on the original EEG signal.
6. The authentication method according to claim 1 or 5, characterized in that, The continuous feature vector is subjected to adaptive differential binarization to obtain the binary brainprint biometrics, wherein the adaptive differential binarization adopts a misalignment comparison and cyclic filling mechanism.
7. The authentication method according to claim 1, characterized in that, The registered error correction codeword and the binary brainprint biometric feature are XORed to obtain the fuzzy commitment value; and the binary brainprint biometric feature of the EEG signal to be tested is XORed with the fuzzy commitment value to obtain the sequence to be decoded.
8. The authentication method according to claim 1, characterized in that, The binary brainprint biometric features are rearranged using a preset permutation sequence. Then, the registered error correction codeword is XORed with the rearranged binary brainprint biometric features to obtain a fuzzy commitment value. Furthermore, the binary brainprint biometric features of the EEG signal to be tested are rearranged in the same way using the preset permutation sequence. Then, the binary brainprint biometric features of the rearranged EEG signal to be tested are XORed with the fuzzy commitment value to obtain the sequence to be decoded.
9. The authentication method according to claim 1, characterized in that, The raw EEG information is sequentially downsampled, filtered, analyzed by independent component analysis, and subjected to common average reference to remove noise and obtain a denoised EEG signal. The time-frequency feature tensor is obtained by performing wavelet scattering transform on the denoised EEG signal.
10. A device for fuzzy commitment authentication using electroencephalogram (EEG) signals, characterized in that, It includes a preprocessing module, a feature extraction module, a binary biometrics acquisition module, a key generation module, an error correction codeword generation module, a key hash acquisition module, a fuzzy commitment value acquisition module, an authentication module, and a storage module; The preprocessing module is used to denoise the acquired raw EEG signals to obtain denoised EEG signals. The feature extraction module includes a time-frequency feature tensor acquisition module and a depth feature acquisition module; wherein, the time-frequency feature tensor acquisition module is used to obtain a time-frequency feature tensor based on the denoised EEG signal, and the depth feature acquisition module is used to obtain a continuous feature vector containing human identity information based on the time-frequency feature tensor; The binary biometric acquisition module is used to obtain binary brain pattern biometrics based on the continuous feature vector; The key generation module is used to generate high-entropy true random keys; The error correction codeword generation module is used to encode the high-entropy true random key into a registered error correction codeword; The key hash acquisition module is used to perform hash calculation on the high-entropy true random key to obtain the registration key hash; The fuzzy commitment value acquisition module is used to generate fuzzy commitment values based on the registered error correction codewords and binary brainprint biometric features; The authentication module is used to acquire the binary brainprint biometric features of the EEG signal to be tested, acquire the sequence to be decoded based on the binary brainprint biometric features and fuzzy commitment value of the EEG signal to be tested, acquire the authentication error correction codeword through the neural LDPC decoder based on the sequence to be decoded, acquire the authentication random key and the authentication key hash corresponding to the authentication random key based on the authentication error correction codeword, and acquire the authentication key hash based on the authentication key hash and the registration key hash to achieve authentication of the EEG signal to be tested; wherein, the neural LDPC decoder contains multiple iterative layers, each iterative layer follows the Tanner graph of the LDPC code, and learnable parameters are set in the verification node layer and / or variable node layer; The storage module is used to store the registration key hash and the fuzzy commitment value.