A system and method for multi-modal data perception and research in a secret space

Abstract

This application relates to the technical field of information security protection, and discloses a system and method for sensing and analyzing multimodal data in classified spaces. The system includes a multimodal module that simultaneously acquires a first vibration signal, a second vibration signal, an acoustic reference signal, and broadband electromagnetic spectrum data; a vibration source module that maps the time-frequency representations of the three signals to construct a coupled tensor, performs block term decomposition under regularization constraints to separate the time-domain components of multiple excitation sources, and identifies laser eavesdropping sources through a transfer metric and outputs a first threat indicator; an electromagnetic detection module that constructs a spatial covariance matrix from the broadband electromagnetic spectrum data to generate a current persistent graph, calculates the topological drift between the current persistent graph and the historical persistent graph, and outputs a second threat indicator after change point detection; and a fusion analysis module that combines evidence of the two threat indicators under time alignment, generates comprehensive alarm information based on spatiotemporal rules, and achieves synchronous sensing and collaborative analysis of laser eavesdropping and unknown electromagnetic espionage threats.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security protection, and discloses a multi-modal data sensing and judgment system for secret-related spaces. BACKGROUND

[0002] In the security protection of secret-related spaces, laser eavesdropping and electromagnetic eavesdropping are two hidden threats that need to be dealt with. The existing technology still has many deficiencies in the actual confrontation environment. For example, in the detection of laser eavesdropping, the existing system collects the vibration signals of the window glass and judges whether there is eavesdropping through an energy threshold or linear coherence between signals. However, when the attacker simultaneously applies an interference signal carrying a fake coherent feature to the glass, the vibration caused by the real eavesdropping, the interference vibration and the vibration caused by the normal indoor voice are mixed in the glass medium, and there is a lack of means to effectively separate multiple excitation sources mixed on the same physical carrier. The attacker's fake interference signal can simulate the surface statistical coherence, but it cannot replicate the causal direction. The existing linear analysis method cannot take advantage of this difference, resulting in high false alarm rate and high missed detection rate. In the detection of electromagnetic eavesdropping, the eavesdropping device uses frequency spectrum agility and noise emission, and the signal is completely hidden in the normal communication frequency band and the non-stationary environment noise. It does not have fixed cyclostationary characteristics or known modulation fingerprints. The frequency spectrum monitoring means relies on energy detection or known signal template matching. When there is no prior signal feature library to compare, it is difficult to discover such unknown abnormal signals only by the structural changes of the electromagnetic environment, forming a detection blind area. SUMMARY

[0003] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title. Such simplifications or omissions cannot be used to limit the scope of the present application.

[0004] The specific invention purposes of the present application are as follows: A vibration source separation and identification mechanism is provided to effectively separate multiple excitation sources mixed on the glass, and to distinguish the real eavesdropping source from the fake interference through the causal direction and time delay characteristics, solving the problem that the existing method is easily deceived by false coherence.

[0005] An electromagnetic environment anomaly detection mechanism is provided to discover hidden eavesdropping signals without relying on a prior signal library by calculating the topological drift of the current and historical persistent graph, solving the problem that unknown frequency spectrum agility signals cannot be sensed.

[0006] On the basis of the above two-channel detection, time alignment evidence fusion is performed to reduce the false alarm rate of a single channel and improve the comprehensive alarm reliability.

[0007] To address the aforementioned technical issues, this application provides a system for sensing and analyzing multimodal data in classified spaces.

[0008] On the one hand, this application provides a system for sensing and analyzing multimodal data in classified spaces, including...

[0009] The multimodal module is used to simultaneously acquire the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data; The vibration source module maps the time-frequency representation of the first vibration signal, the second vibration signal, and the acoustic reference signal, constructs a coupling tensor, applies a regularization constraint to the coupling tensor, and performs block term decomposition to separate the time-domain components of multiple excitation sources. Through the transmission metric between the excitation source and the first vibration signal, the laser eavesdropping source is identified and the first threat indicator is output. The electromagnetic detection module constructs a spatial covariance matrix based on the time-frequency blocks of the broadband electromagnetic spectrum data to generate the current persistent graph, and calculates the distance between the current persistent graph and the historical persistent graph as the topology drift. The change point detects the topology drift and outputs the second threat indicator. The fusion analysis module combines the first threat indicator and the second threat indicator with evidence aligned to time, and generates comprehensive alarm information according to preset spatiotemporal rules.

[0010] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: The first vibration signal was collected by a non-contact optical vibration measuring device pointing towards the window glass; The second vibration signal was acquired by a contact vibration sensor attached to the same glass. The acoustic reference signal is acquired by an indoor microphone; The broadband electromagnetic spectrum data is collected by multiple broadband receiving antennas.

[0011] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: Time-frequency transformations were performed on the first vibration signal, the second vibration signal, and the acoustic reference signal to obtain time-frequency spectra. Each time-frequency spectrum is mapped to the regenerating kernel Hilbert space by a nonlinear kernel function to obtain the high-order nonlinear similarity between each modal signal that cannot be linearly expressed in the original time-frequency domain. In the regenerative kernel Hilbert space, the higher-order cross-cumulants of the first vibration signal, the second vibration signal, and the acoustic reference signal are calculated, and a coupling tensor is constructed such that the elements of the coupling tensor represent the statistical correlation between different modes at different time and frequency points.

[0012] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: The method for applying regularization constraints and performing block decomposition to the vibration source module is as follows: The regularization constraint originates from a fractional-order model characterizing the damping properties of the target glass, which includes a fractional-order time derivative term to represent the viscoelastic memory of the classified space glass. Using the fractional-order model, a physical regularization term is constructed, which penalizes deviations from the fractional-order differential relationship that each separated excitation source should satisfy between the first vibration signal, the second vibration signal, and the acoustic reference signal. During the block decomposition process, the coupling tensor is represented as the sum of multiple low-rank tensor blocks, each tensor block corresponding to an independent excitation source, and the physical regularization term is introduced into the decomposition objective function so that the separation result is forced to follow the fractional order law of the glass. By solving the block decomposition with physical regularization constraints, multiple independent time-domain components of the excitation source that physically conform to the glass vibration transmission characteristics are obtained.

[0013] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: The method for identifying the laser eavesdropping source and outputting the first threat indicator by measuring the transmission of the excitation source and the first vibration signal is as follows: The transfer measure includes nonlinear transfer entropy and symbol transfer entropy based on sorting entropy; Calculate the information transmission between each separated excitation source and the first vibration signal to determine whether there is a causal drive from the first vibration signal to the excitation source, and estimate the time delay of information transmission; An excitation source that simultaneously satisfies the causal direction of the first vibration signal pointing to the excitation source and the time delay matching the preset delay range of the laser eavesdropping process is marked as a real laser eavesdropping source, and the directional information flow intensity and information transmission time delay of the real laser eavesdropping source are obtained. The first threat indicator is generated by measuring the strength of the directional information stream and the stability of its delay, which are identified as real laser eavesdropping sources.

[0014] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: The electromagnetic detection module constructs a spatial covariance matrix based on the time-frequency blocks of the broadband electromagnetic spectrum data, obtains multiple positive definite matrices and embeds them into a positive definite manifold, constructs a filtered simple complex based on geodesic distance on the positive definite manifold and calculates the persistent homology, and generates the current persistent graph. The broadband electromagnetic spectrum data is divided into several time-frequency blocks according to time frames and frequency sub-bands; For the multi-channel complex baseband sampling data within each time-frequency block, calculate its sample covariance matrix; Structural constraints are applied to the sample covariance matrix, and it is modified into a positive definite matrix with a symmetric structure, which serves as the spatial covariance matrix of the time-frequency block.

[0015] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: Assign an affine-invariant Riemannian metric to each of the spatial covariance matrices, such that all the positive definite matrices to be processed form a positive definite matrix manifold with an affine-invariant Riemannian metric. Using geodesic distance as the distance metric between two points on the positive definite matrix manifold, a filtered simplex is constructed. As the distance threshold gradually increases, edges, triangles, and higher-order simplexes are generated sequentially between points on the manifold whose distance is less than the current threshold, forming a simplex filtering sequence that evolves with the threshold. The continuous homology of the filtering sequence is calculated, the occurrence and extinction times of each topological feature during the filtering process are recorded, and a persistent barcode and persistent map labeled with birth and death parameter pairs are generated. The persistent graph of the current monitoring time window is used as the current persistent graph, and each point of the current persistent graph represents a stable generator in the electromagnetic environment topology.

[0016] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: A collection of persistent graphs acquired multiple times under normal electromagnetic environments throughout history was constructed. Using the Wasserstein distance between persistent graphs as a metric, the topological difference between the current persistent graph and the set of historical persistent graphs is calculated to obtain a sequence of topological drift over time. Perform Bayesian online change point detection on the topological drift sequence and estimate the posterior probability of the current time being an anomalous change point in real time; When the posterior probability exceeds a preset confidence threshold, it is determined that a topological anomaly caused by an unknown eavesdropping signal has occurred in the electromagnetic environment, and the current posterior probability is output as the second threat indicator.

[0017] As a preferred embodiment of the multimodal data sensing and analysis system for classified spaces in this application, wherein: The first threat indicator and the second threat indicator are respectively converted into probability assignment functions under the same identification framework; By employing the evidence combination rule, the first basic probability assignment and the second basic probability assignment that are aligned in time are fused to obtain the joint confidence level allocation; The joint confidence level allocation is adjusted according to a preset spatiotemporal logic rule, which includes: when the laser eavesdropping event represented by the first threat indicator and the electromagnetic anomaly event represented by the second threat indicator overlap within a time window, the confidence level of the eavesdropping attack is increased; when only the second threat indicator appears and the first threat indicator does not simultaneously indicate laser eavesdropping, the confidence level of the eavesdropping attack is reduced and marked as an event to be observed.

[0018] As a preferred embodiment of the method for sensing and analyzing multimodal data in classified spaces as proposed in this application, wherein: S1 simultaneously acquires the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data; S2 processes the first vibration signal, the second vibration signal, and the acoustic reference signal to construct a higher-order coupling tensor. It applies physical regularization constraints to the coupling tensor and performs block term decomposition to separate the time-domain components of multiple excitation sources. Based on the nonlinear directional information transmission measure between each excitation source and the first vibration signal, it identifies laser eavesdropping sources and generates a first threat index. S3 processes the broadband electromagnetic spectrum data to construct a spatial covariance matrix and embed it into a positive definite manifold. A filtered complex is constructed on the manifold, and a continuous homology is calculated to generate the current persistent graph. The topological drift between the current persistent graph and the historical persistent graph is calculated, and a second threat index is generated through change point detection. S4 performs time-aligned evidence fusion on the first threat indicator and the second threat indicator to generate comprehensive alarm information.

[0019] The beneficial effects of this application are as follows: This application constructs a coupled tensor from three signals—the first vibration signal, the second vibration signal, and the acoustic reference signal—using a vibration source module, and performs block term decomposition under regularization constraints. This achieves effective separation of multiple excitation sources under nonlinear aliasing conditions, overcoming the problem that traditional methods cannot simultaneously separate multiple excitation sources.

[0020] This application identifies laser eavesdropping sources based on the transmission measure between the excitation source and the first vibration signal using a vibration source module. It distinguishes between real eavesdropping sources and fake coherent interference by utilizing causal direction and time delay characteristics, thus solving the problem that traditional coherent analysis is easily deceived by false interference.

[0021] This application uses an electromagnetic detection module to construct a spatial covariance matrix based on time-frequency blocks of broadband electromagnetic spectrum data to generate the current persistent graph, and calculates the topological drift between the current persistent graph and the historical persistent graph. It can detect anomalies based solely on minute changes in the topology of the electromagnetic environment without a prior signal feature library, thus solving the problem that spectrum-agile eavesdropping signals cannot be detected by traditional template matching or energy detection.

[0022] This application combines evidence from the first and second threat indicators in a time-aligned manner through a fusion analysis module, and generates a comprehensive alarm based on spatiotemporal rules. It also coordinates and verifies the detection results of the optical vibration channel and the electromagnetic channel, effectively reducing the false alarm rate when a single channel is judged independently, and improving the reliability of the system's comprehensive analysis in adversarial environments. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained through these drawings without creative effort. Wherein: Figure 1 This application provides an overall architecture and data flow diagram for a multimodal data perception and analysis system for classified spaces; Figure 2 This application provides a flowchart of the vibration source module processing in a multimodal data sensing and analysis system for classified spaces. Figure 3 This application provides a flowchart of the topology drift detection process for an electromagnetic detection module in a classified space multimodal data sensing and analysis system. Figure 4 This application provides a decision logic diagram for a fusion analysis module in a multimodal data perception and analysis system for classified spaces. Figure 5 This application provides a multimodal collaborative temporal interaction diagram for use in a multimodal data perception and analysis system for classified spaces. Detailed Implementation

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments. Example

[0027] This embodiment provides a system for sensing and analyzing multimodal data in classified spaces, including: The multimodal module is used to simultaneously acquire the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data; In this application, a preferred implementation method for a multimodal module includes: The first vibration signal is collected by a non-contact optical vibration measuring device pointing towards the window glass. The optical vibration measuring device can be a laser Doppler vibrometer, which emits a detection laser beam towards the window glass of the confidential space, receives the laser reflected back from the glass surface, and outputs an analog or digital signal reflecting the vibration velocity or displacement of the glass surface by detecting the Doppler frequency shift or phase change of the reflected light.

[0028] The second vibration signal is acquired by a contact vibration sensor attached to the same glass. The contact vibration sensor can be a piezoelectric accelerometer or a piezoelectric thin film sensor, which is attached to the inner surface of the glass by a coupling agent or mechanical clamp. Its installation position is close to or coaxial with the measurement spot position of the optical vibration measuring device, so as to obtain the vibration response of the same glass area under the same excitation conditions.

[0029] The acoustic reference signal is collected by an indoor microphone installed at a predetermined location within the confidential space. The microphone is used to pick up the indoor sound field, record acoustic events such as normal indoor speech and environmental noise, and provide reference information at the sound source end for vibration source separation.

[0030] Broadband electromagnetic spectrum data is acquired by multiple broadband receiving antennas. The broadband receiving antennas cover the predetermined monitoring frequency band. The multiple antennas can be arranged in a spatial diversity configuration. The signals received by each antenna are synchronously down-converted and analog-to-digital converted by a multi-channel receiver, outputting multiple complex baseband sampling data streams.

[0031] like Figure 2 As shown, the vibration source module maps the time-frequency representation of the first vibration signal, the second vibration signal, and the acoustic reference signal, constructs a coupling tensor, applies physical regularization constraints based on a fractional-order dynamic model characterizing the damping properties of the target glass to the coupling tensor, and performs block term decomposition to separate the time-domain components of multiple excitation sources. Through the transmission metric between the excitation source and the first vibration signal, the laser eavesdropping source is identified and the first threat indicator is output. After acquiring the first vibration signal, the second vibration signal, and the acoustic reference signal, the vibration source module performs time-frequency transformation on the three signals to obtain their respective time-frequency spectra. The time-frequency transformation can be a short-time Fourier transform, which maps each signal from the time domain to a two-dimensional time-frequency plane, so that the components of each excitation source that are mixed in the time domain exhibit different energy distribution characteristics in the time-frequency domain.

[0032] After obtaining the three time-frequency spectra, each time-frequency spectra is mapped to the regenerating kernel Hilbert space using a nonlinear kernel function. The nonlinear kernel function enables the linearly inseparable signal features in the original time-frequency domain to be represented in the form of an inner product in the regenerating kernel Hilbert space. This obtains the nonlinear similarity relationships between different modal signals that cannot be directly expressed by linear operations in the original time-frequency domain. After mapping, the time-frequency structure of each modal signal is embedded in a high-dimensional feature space, preserving richer nonlinear interaction information than the original domain.

[0033] In the regenerating kernel Hilbert space, the higher-order cross-cumulants of the first vibration signal, the second vibration signal, and the acoustic reference signal are calculated. The higher-order cross-cumulants are multivariate statistics used to capture the higher-order statistical dependencies beyond the second-order correlation between the three signals. In the scenario of multiple excitation sources aliasing, the sources cannot be effectively distinguished at the second-order statistical level, but there are differences in the higher-order statistical characteristics. The calculation of the higher-order cross-cumulants preserves the differences.

[0034] A coupling tensor is constructed using the calculated higher-order cross-cumulants. Each element of the coupling tensor corresponds to the higher-order cross-cumulant values ​​of different modal signals at different time-frequency points and with different time-delay combinations, so that each element of the coupling tensor represents the statistical correlation between the first vibration signal, the second vibration signal, and the acoustic reference signal on a specific time-frequency coordinate. This coupling tensor is organized in the form of a multidimensional array, fully preserving the higher-order statistical structure formed by the aliasing excitation sources among the multimodal signals.

[0035] In this application, an example of a preferred vibration source module for constructing coupled tensors and applying constraint decomposition includes: For the first vibration signal Second vibration signal Acoustic reference signal Perform short-time Fourier transforms on each component to obtain the time-frequency spectra. , and , where τ is the time frame index and f is the frequency point.

[0036] Each time-frequency point is mapped to the reproducing kernel Hilbert space using a nonlinear kernel function ϕ(*), and the mapping satisfies... , where κ(*,*) is the selected kernel function.

[0037] The above cross-cumulants are organized into a fourth-order tensor T based on the time-frequency points and time delays of the three signals. Its elements are: ;

[0038] Coupled Tensor Each dimension corresponds to the time-frequency coordinates and time delay of different modes, preserving the nonlinear high-order statistical correlation structure between multimodal signals. Indices i, j, k traverse the time-frequency grid points of the first vibration signal, the second vibration signal, and the acoustic reference signal, respectively. Cum(*,*) is a third-order joint cumulant operator. Each element of the third-order tensor characterizes the nonlinear high-order correlation strength of the three signals under a specific combination of time-frequency coordinates.

[0039] When applying regularization constraints and performing block term decomposition, a pre-built fractional-order dynamic model is used.

[0040] Where W is the glass vibration displacement. The fractional derivative operator, 0 < α < 1, p is the density coefficient, c is the identified damping coefficient, and μ is the stiffness parameter. The waveform projected onto the first vibration signal, the second vibration signal, and the acoustic reference signal for the r-th excitation source is... , , 1, 2, and 3 are channel identifiers, and physical regularization terms are constructed.

[0041] The regularization term is used to measure the degree of deviation between the separated vibration source and the sound source from satisfying the fractional vibration equation.

[0042] Furthermore, let the acoustic excitation force acting on the glass be... Then the fractional equation should be: ;

[0043] The corrected physical regularization term is: ;

[0044] Where β is the sound pressure conversion coefficient, which can be taken as β=5. N / Pa, the unit of measurement, represents the magnitude of the equivalent force produced by a unit sound pressure acting on a unit area of ​​glass surface. Its value can be obtained through offline calibration experiments under standard sound source excitation. Norm We take the L2 norm, which is the vector 2 norm in the actual discrete implementation.

[0045] An example of a preferred fractional derivative discrete approximation: ;

[0046] Where h is the sampling interval, and N = [t / h] is the truncation window length. The coefficients are the generalized binomial coefficients. This is a gamma function.

[0047] The objective function for block decomposition is: ; in Let R be the r-th tensor block, R be the number of excitation sources, and λ be the trade-off parameter. By iteratively minimizing the objective function using the alternating direction multiplier method, several low-rank tensor blocks are obtained. Each tensor block corresponds to a time-domain component of an excitation source that is statistically independent and conforms to the glass vibration transmission characteristics in terms of physical waveform.

[0048] The method for applying regularization constraints and performing block decomposition to the vibration source module is as follows: The regularization constraint originates from a fractional-order model characterizing the damping properties of the target glass, which includes a fractional-order time derivative term to represent the viscoelastic memory of the classified space glass. Using the fractional-order model, a physical regularization term is constructed, which penalizes deviations from the fractional-order differential relationship that each separated excitation source should satisfy between the first vibration signal, the second vibration signal, and the acoustic reference signal. During the block decomposition process, the coupling tensor is represented as the sum of multiple low-rank tensor blocks, each tensor block corresponding to an independent excitation source, and the physical regularization term is introduced into the decomposition objective function so that the separation result is forced to follow the fractional order law of the glass. By solving the block decomposition with physical regularization constraints, multiple independent time-domain components of the excitation source that physically conform to the glass vibration transmission characteristics are obtained.

[0049] In this application, a preferred method for applying regularization constraints and performing block term decomposition on a vibration source module includes: First, a fractional-order model characterizing the damping properties of the target glass is established. The glass in the classified space window is a viscoelastic material, and its micro-vibration response under external excitation depends not only on the current excitation but also on the continuous influence of historical excitations, exhibiting a memory effect. The memory-like damping behavior is obtained in a compact form using a fractional-order derivative operator, which is expressed as a differential equation containing a fractional-order time derivative: ; Where W represents the glass vibration displacement, Here, f(t) is an α-order fractional derivative operator (0 < α < 1), and f(t) is the total excitation force acting on the glass. Parameters α, c, and k are obtained through offline impact tests to systematically identify the target glass and are used to accurately reflect the actual viscoelastic mechanical properties of the glass.

[0050] It should be noted that α can be chosen to be ∈(0.3,0.7), preferably 0.5; ρ, c, and k can be identified through offline tapping tests;

[0051] An example of a preferred fractional-order model for constructing physical regularization terms: In a real physical process, the projected waveforms of each excitation source on the first vibration signal, the second vibration signal, and the acoustic reference signal satisfy the fractional-order dynamic relationship of the glass.

[0052] For example, The source signal of the r-th independent excitation source is defined as having projection components in each channel after passing through different physical paths. If a certain excitation source... When applied to glass, the physical vibration displacement generated on the glass surface is denoted as . Specifically, if the vibration displacement generated on the glass surface by a certain excitation source is... The response of the vibration in the optical vibration meter channel and There is a differential relationship between them, such as If the incentive source Part of the sound originates from the indoor sound field, and its projection in the acoustic channel is denoted as... Then acoustic projection With glass vibration projection The input-output relationship should satisfy the fractional-order vibration equation characterizing the glass damping properties, when a certain candidate source component is given. The corresponding projection components on the three signals { If the above physical relationship is not satisfied, then the candidate source component is determined. It contains separation errors or is composed of mismatched interference sources.

[0053] The physical regularization term can be constructed as a measure of the residual between the separated vibration source waveform and the sound source waveform with respect to the fractional-order vibration equation. For example, for the r-th excitation source, the physical regularization term can be expressed as a calculation expression for the physical regularization term.

[0054] The regularization term calculates the deviation between the acoustic response predicted by the fractional-order model and the actual separated acoustic component in the separation result of the vibration component. The larger the deviation, the heavier the penalty.

[0055] Embedding the physical regularization term into the objective function of tensor block decomposition, tensor block decomposition represents the coupled tensor T as R low-rank tensor blocks. The sum of these, each tensor block corresponds to an independent excitation source.

[0056] It is important to note that the objective function for separation considers both the sum of each tensor block and the reconstruction error of the original coupled tensor, using fidelity to measure the completeness of the separation. The physical regularization terms of all excitation sources are weighted to ensure the physical rationality of the separation result. The objective function is iteratively minimized using the alternating direction multiplier method. During the iteration, each tensor block is updated step by step until the data fidelity term and the physical regularization term reach a balance.

[0057] After the solution is completed, the time-domain component extracted from each tensor block is the contribution of an independent excitation source to the first vibration signal, the second vibration signal, and the acoustic reference signal. Since the decomposition process is constrained by the fractional-order physical model, the separated excitation sources are not only statistically independent, but their time-domain waveforms also conform to the real vibration transmission characteristics of the glass medium, which enables the vibration source caused by real laser eavesdropping to be effectively distinguished from the interference source or normal speech source at the physical level.

[0058] The method for identifying the laser eavesdropping source and outputting the first threat indicator based on the transmission metric between the excitation source and the first vibration signal is as follows: Indoor voice communication causes micro-vibrations in the window glass. An attacker emits a laser from a distance towards the glass and receives the reflected light. By demodulating the phase or frequency changes of the reflected light, the indoor voice communication is reconstructed. The first vibration signal is collected by an optical vibration meter, reflecting the result of optical demodulation of the glass vibration. For the glass vibration caused by indoor voice communication, the acoustic signal drives both the vibration and optical signals. However, for a real laser eavesdropping attack, the attacker actively emits a laser and receives the reflected light. The eavesdropping behavior itself does not generate additional mechanical drive on the glass. This application uses an optical vibration meter to actively emit and receive a detection laser. When the attacker's eavesdropping laser and the detection laser simultaneously illuminate the same glass area, optical interference may occur between the reflected light of the two lasers, introducing an additional component caused by the eavesdropping laser into the demodulation output of the optical vibration meter.

[0059] The transfer measure includes nonlinear transfer entropy and symbol transfer entropy based on sorting entropy; Calculate the information transmission between each separated excitation source and the first vibration signal to determine whether there is a causal drive from the first vibration signal to the excitation source, and estimate the time delay of information transmission; Specifically, for each separated excitation source component, its time-domain waveform in the first vibration channel and the second vibration channel is extracted respectively. The nonlinear transfer entropy and symbol transfer entropy between the first vibration signal and each channel component of the excitation source are calculated. The level of information transmission is determined by permutation test. If a directional information flow from the direction of the first vibration signal to the excitation source is detected, it indicates that there is a causal driving relationship between the excitation source and the optical channel, which is dominated by or manifested by the optical signal. This is consistent with the causal direction characteristics formed between the optical and vibration channels by laser eavesdropping behavior. The time delay parameter of this causal driving can be obtained by estimating the embedding delay during the transfer entropy calculation process.

[0060] If the excitation source simultaneously satisfies the causal direction of the first vibration signal pointing to the excitation source, and the estimated time delay parameter falls within the preset delay range of the laser eavesdropping physical process, then the excitation source is marked as a real laser eavesdropping source.

[0061] For vibration sources caused by normal indoor speech, the causal direction is that the acoustic signal drives the vibration signal, rather than the first vibration signal pointing to the excitation source. Therefore, they will not be mislabeled. For environmental pseudo-vibration sources, there is no directional information flow between them and the first vibration signal, so they are naturally excluded.

[0062] An excitation source that simultaneously satisfies the causal direction of the first vibration signal pointing to the excitation source and the time delay matching the preset delay range of the laser eavesdropping process is marked as a real laser eavesdropping source, and the directional information flow intensity and information transmission time delay of the real laser eavesdropping source are obtained. The first threat indicator is generated by measuring the strength of the directional information stream and the stability of its delay, which are identified as real laser eavesdropping sources.

[0063] In this application, a preferred method for generating a first threat indicator includes: For each separated excitation source, the time-domain waveform of the second vibration signal channel and the corresponding time-domain waveform of the first vibration signal channel are extracted. Since the three signals are sampled synchronously and share the time-domain structure during block decomposition, the source components of each channel are aligned in time.

[0064] Furthermore, the nonlinear transfer entropy between the first vibration signal and the excitation source component is calculated separately. The calculation steps of the nonlinear transfer entropy include: reconstructing the state space of the two time series respectively, and constructing a vector sequence using a delayed embedding method; in each reconstructed state, estimating the conditional probability density using a neighborhood point statistical method, and thereby calculating the directional information transfer amount from the first vibration signal to the excitation source component. The information transfer amount is used to measure the reduction in uncertainty of the current state of the excitation source due to the historical state of the first vibration signal. The directionality determines the causal orientation, and the magnitude reflects the strength of the causal drive. By changing the delay parameter, the embedding delay corresponding to the maximum value of information transfer can be estimated. The embedding delay is the time delay estimate of the causal drive.

[0065] To further enhance the ability to capture complex dynamics, symbolic transfer entropy based on sorting entropy is also calculated. Specifically, the two sequences are converted into sorted symbol sequences, that is, symbol encoding is performed according to the arrangement pattern of numerical values ​​in the time window. Then, the transition probability between different symbols is calculated, and the transfer entropy is obtained. The dynamic capture method disclosed in this application, which calculates the transfer entropy, is not affected by the instantaneous fluctuation of the signal amplitude and can more robustly reflect the causal transmission of the internal temporal structure of the signal. It is applicable to glass vibration systems with memory effects.

[0066] To determine the statistical significance of information transmission, a permutation test was performed on the above-mentioned transmission entropy calculation results: by randomly shuffling the time sequence of the first vibration signal multiple times and recalculating the transmission entropy, the information transmission distribution under the null hypothesis was obtained. When the original transmission entropy value exceeded the standard level threshold, it was determined that there was a causal drive from the first vibration signal to the excitation source.

[0067] Because the coherent interference between the eavesdropping laser and the detection laser forms a rapidly decoherent non-stationary noise source in the optical path, this interference noise manifests as a transient change in phase noise in the optical vibration measurement channel. When the interference component is separated into an independent excitation source through tensor decomposition, there is a unidirectional information transfer between it and the residual component of the first vibration signal: the random fluctuations of the interference source will lead its weak coupling response in the glass vibration, i.e., the second channel, in time. Therefore, in the calculation of transfer entropy, it presents a statistical causal direction from the first vibration signal to the excitation source. A preferred method for calculating transfer entropy includes: ; After completing the causal determination and time delay estimation, each excitation source is screened: the excitation source that meets the following two conditions is marked as the real laser eavesdropping source: the first condition is that the causal direction is from the first vibration signal to the excitation source; the second condition is that the estimated time delay falls within the preset delay range of the laser eavesdropping process.

[0068] It should be noted that X and Y are the first vibration signal and the excitation source component, respectively. Let Y be the state at time t+1. Let Y be the m-dimensional delayed embedding vector of the sequence. Let X be the n-dimensional delayed embedding vector of the sequence. To address the embedding delay, the embedding dimensions m and n can be determined using the spurious nearest neighbor method. The summation iterates through all possible state transition combinations, and the conditional probability p(*|*) is estimated from the samples using a neighborhood estimation method.

[0069] The preset range is determined comprehensively based on physical parameters such as the optical path distance between the optical vibration measuring device and the glass, the signal transmission delay, and the glass vibration propagation time. A preferred example is a time interval. For source components that do not meet the conditions, such as those whose causal direction is the excitation source driving the first vibration signal, or whose time delay exceeds a reasonable range, they are determined to be indoor voice, environmental pseudo-vibration, or other irrelevant sources and are not included in the eavesdropping alarm category.

[0070] In eliminating interference laser signals forged by attackers, the fractional-order memory effect of glass was utilized. In real laser eavesdropping, the optical anomaly component associated with the eavesdropping laser is formed through the actual glass reflection path. The relationship between this component and the vibration channel includes the viscoelastic memory effect of the target glass. Although the interference laser released by the attacker can be statistically correlated with the first vibration signal, because it does not pass through the actual transmission path of the target glass, its corresponding transmission measures, including the time-varying curve of the transmission entropy and the distribution of the symbol transfer entropy, lack the continuous fluctuation characteristics and slow decay trend unique to the fractional-order memory effect.

[0071] By performing fractal feature analysis on the time series of the transfer entropy of each source component, or by comparing the degree of deviation of its statistical distribution from the normal reference pattern, sources that do not possess this memory feature can be excluded from the candidates for eavesdropping sources.

[0072] After marking is completed, for the component identified as a real laser eavesdropping source, its transmission entropy value is obtained as a measure of the intensity of the directional information flow, and the variance of the source delay estimate under multiple consecutive time windows is calculated as a measure of the delay stability.

[0073] The higher the intensity of the directional information flow, the tighter the causal coupling; the more stable the time delay, the higher the credibility of the causal relationship originating from a fixed physical propagation path.

[0074] The information flow intensity and latency stability are weighted and combined to generate a scalar value, which is then output as the primary threat indicator.

[0075] like Figure 3 As shown, the electromagnetic detection module constructs a spatial covariance matrix based on the time-frequency blocks of the broadband electromagnetic spectrum data, obtains multiple positive definite matrices and embeds them into a positive definite manifold, constructs a filtered simple complex based on the geodesic distance on the positive definite manifold and calculates the persistent homology, generates the current persistent graph, and calculates the distance between the current persistent graph and the historical persistent graph as the topological drift. The change point detects the topological drift and outputs a second threat indicator characterizing electromagnetic spatial anomalies. In this application, a preferred method for constructing a spatial covariance matrix using an electromagnetic detection module includes: The electromagnetic detection module receives complex baseband sampling data streams from multiple broadband receiving antennas. First, it divides the continuous time-spectrum data into multiple time-frequency blocks according to the time frame length and frequency sub-band width. The division of frequency sub-bands is based on the frequency band coverage of the broadband receiving antennas and the frequency domain resolution requirements of the covariance matrix estimation. Each time-frequency block contains complex baseband sampling data from multiple antenna channels in the same time frame and the same frequency sub-band, forming a multi-channel complex data matrix. The number of rows is the number of antenna channels, and the number of columns is the number of sampling points in the time-frequency block.

[0076] It is important to note that the selection of time frame length must take into account both the ability to capture burst signals and the constraints of computing resources.

[0077] For each time-frequency block, the electromagnetic detection module calculates its sample covariance matrix. Each diagonal element of the sample covariance matrix represents the signal power of the corresponding antenna channel in that time-frequency block, and each off-diagonal element represents the complex correlation coefficient between two different antenna channels. The complex correlation coefficient contains both amplitude correlation and phase difference information.

[0078] The covariance matrix is ​​used to characterize the spatial statistical features of multi-channel signals within a corresponding time-frequency block, namely the energy distribution of electromagnetic signals at different spatial locations and their spatial coherence structure.

[0079] It should be noted that, under ideal conditions, the sample covariance matrix should be a positive definite matrix. In actual calculations, due to the estimation error introduced by the finite sampling length and the influence of receiving channel noise, the sample covariance matrix may be numerically rank deficient or semi-positive definite, which does not meet the requirement of positive definiteness of matrix in Riemannian manifold calculations. Therefore, after calculating the sample covariance matrix for each time-frequency block, the electromagnetic detection module further applies structural constraints and makes corrections to obtain a matrix with strict positive definiteness.

[0080] One preferred implementation of the structural constraint is the Toeplitz structural constraint.

[0081] Specifically, based on the geometric arrangement of broadband receiving antenna arrays, when the antenna array is a uniform linear array or can be approximated as a uniform linear array, under the far-field and narrow-band assumptions, the spatial correlation coefficient between different antenna channels is related to the antenna spacing but not to the antenna index.

[0082] The elements on each diagonal of the spatial covariance matrix are equal. The electromagnetic detection module performs Toeplitz correction on the sample covariance matrix by ensuring that the elements on each diagonal of the spatial covariance matrix are equal: the average of the elements on each diagonal is taken as the uniform value of all elements on the diagonal. The corrected matrix retains the main spatial statistical characteristics of the original sample covariance matrix and also satisfies the requirements of Hermite symmetry and Toeplitz structure.

[0083] After Toeplitz correction, the electromagnetic detection module further ensures the positive definiteness of the matrix through regularization. It performs spectral correction on the matrix after Toeplitz structure correction by replacing non-positive values ​​in its eigenvalues ​​with a small positive constant to ensure that the matrix is ​​strictly positive definite while minimizing the impact on the original matrix structure.

[0084] The electromagnetic detection module generates a positive definite matrix for each time-frequency block. The positive definite matrices corresponding to all time-frequency blocks together form a matrix set. While reflecting the spatial statistical characteristics of the electromagnetic signal in the corresponding time-frequency block, the positive definiteness of each positive definite matrix provides basic data for embedding a positive definite matrix manifold equipped with affine invariant Riemannian metric, calculating geodesic distances, and constructing filtered simple complexes.

[0085] An example of calculating geodesic distance using a preferred affine-invariant Riemannian metric is as follows: ; Where log(⋅) represents the logarithm of the matrix, i.e., the eigenvalue decomposition of the matrix. After that, take diagonal matrix Each element is the natural logarithm of its eigenvalues. The Frobenius norm is defined as follows: , Let represent the square root of the inverse of a positive definite matrix A, which can be calculated using Cholesky decomposition. The ground distance satisfies for any invertible matrix W. Affine invariance.

[0086] After obtaining multiple positive definite spatial covariance matrices corresponding to each time-frequency block, the electromagnetic detection module embeds a positive definite manifold, constructs a filtered complex, and generates the current persistent graph. The specific implementation methods include: The positive definite matrices corresponding to all time-frequency blocks constitute a set of positive definite matrices to be processed. The electromagnetic detection module assigns an affine invariant Riemannian metric to each spatial covariance matrix in the set of positive definite matrices. All positive definite matrices constitute a positive definite matrix manifold with the Riemannian metric.

[0087] It is important to note that in the case of affine invariant Riemannian metric, the space of positive definite matrices is not a Euclidean space. The Euclidean distance is not invariant to the joint similarity transformation of matrices. That is, when two positive definite matrices undergo the same invertible matrix similarity transformation, the Euclidean distance will change.

[0088] The metric selection is based on the following: In electromagnetic environment monitoring scenarios, factors such as gradual changes in the gain of the receiving channel and device aging can cause the multi-channel covariance matrix to undergo a similarity scaling transformation. If the distance metric does not have invariance to the overall similarity scaling transformation of the multi-channel covariance matrix, normal environmental fluctuations will be misjudged as abnormal changes in the manifold structure. The affine invariant Riemannian metric maintains the geodesic distance under the joint similarity transformation of invertible matrices. That is, after two positive definite matrices undergo the same invertible matrix similarity transformation, the geodesic distance on the manifold remains unchanged, so that common-mode changes such as receiving channel gain fluctuations are uniformly absorbed. The relative geometric relationship of each matrix point on the manifold only responds to the changes in the spatial structure of the electromagnetic environment, rather than global scaling disturbances, and the deviation of the abnormal matrix from the normal background can be accurately measured.

[0089] On a positive definite matrix manifold, the electromagnetic detection module uses geodesic distance as the distance metric between matrix points. Geodesic distance is the shortest curve length connecting two points on the manifold, used to reflect the true spacing between matrix points under the inherent geometry of the manifold.

[0090] The electromagnetic detection module constructs a filtered Vietoris-Rips simplex complex using geodesic distances. The specific steps are as follows: a set of distance thresholds are preset in ascending order. For the smallest distance threshold, all matrix points on the manifold are traversed, and points whose geodesic distance does not exceed the distance threshold are connected to form edges. The resulting complex consists of several isolated edges and scattered connected components. As the distance threshold gradually increases, previously isolated connected components are merged by adding new edges to connect them. When the geodesic distance between any three points does not exceed the current threshold, triangles are generated with the three points as vertices, and two-dimensional surfaces begin to appear. As the current threshold continues to increase, the addition of more edges and surfaces causes some ring structures to be filled by triangles and disappear. The internal region of the complex is gradually filled with higher-dimensional simplexes. This process continues, with each threshold level corresponding to a simplex complex. The simplex complexes at each level are arranged in ascending order of the threshold, forming a filtered simplex complex sequence.

[0091] The electromagnetic detection module performs continuous cohomology calculation on the filtered simple complex sequence, continuously cohomologically tracking the first appearance threshold and final disappearance threshold of each topological generator in the filtered sequence, which are recorded as the birth value and the extinction value, respectively. With the birth value as the x-axis and the extinction value as the y-axis, each topological generator is marked as a data point on the plane. The data points of all generators together constitute the persistent graph of the current monitoring time window. Points near the diagonal on the persistent graph correspond to transient structures with short lifespans, while points far from the diagonal correspond to topological structures with long durations and stable spatial correlation support in the electromagnetic environment.

[0092] A preferred method for persistent cohomology computation and persistent graph generation is as follows: The electromagnetic detection module performs continuous cohomology calculations on the above filtered simple complex sequence, tracking the lifecycle of each topological feature during the filtering process: the first appearance of each topological feature in a filtered complex at a certain threshold is recorded as the birth time; as the threshold continues to increase, the topological feature is filled or merged with other features in subsequent complexes and disappears, recorded as the extinction time. 0-dimensional continuous cohomology tracks the generation and merging of connected components, and 1-dimensional continuous cohomology tracks the generation and filling of rings. The electromagnetic detection module records the birth parameters and extinction parameters of each topological feature, forming birth and death parameter pairs, and organizes them in two forms of visualization and computation: persistent barcodes and persistent graphs.

[0093] In a persistent graph, each point corresponds to a topological feature. The horizontal axis represents the birth value of the topological feature, and the vertical axis represents the extinction value. Points located above the diagonal indicate that the topological feature lasts for a time interval during the filtering process, that is, from birth to extinction, it spans the threshold range. The farther a point is from the diagonal, the longer the duration of the corresponding topological feature, and the more stable it is in the data. Points closer to the diagonal have shorter durations of the corresponding topological feature, which are instantaneous fluctuations or noise in the data.

[0094] The persistent graph of the current monitoring time window is used as the current persistent graph, and each point of the current persistent graph represents a stable generator in the electromagnetic environment topology.

[0095] During the initial deployment phase or periods confirmed to be in a safe electromagnetic environment, broadband electromagnetic spectrum data is continuously collected, and persistent maps for each monitoring time window are generated. Persistent maps acquired under normal electromagnetic conditions are aggregated to form a historical background persistent map set. This historical background persistent map set is used to represent the typical form of the electromagnetic environment topology of the classified space under conditions without espionage attacks and its normal fluctuation range. Specifically, it is represented by the point cloud distribution pattern formed by the birth and death parameters of each generator on the persistent map. As the system runs for a long time, the historical background persistent map set can be maintained through a sliding window to ensure that it always reflects the baseline of the normal electromagnetic environment at the current stage and avoids baseline obsolescence due to the addition or removal of legitimate wireless devices or seasonal changes in spectrum usage.

[0096] When it is necessary to make an anomaly assessment of the electromagnetic environment at the current moment, the electromagnetic detection module calculates the topological difference between the current persistent graph and the set of historical background persistent graphs. The topological difference is measured by the Wasserstein distance between the persistent graphs. The Wasserstein distance is used to measure the minimum total transmission cost required to match a point in one persistent graph to another.

[0097] The persistent graph is a two-dimensional graph formed by marking each topological feature as a planar data point with the birth parameter of the topological feature as the horizontal axis and the death parameter as the vertical axis.

[0098] In this application, a preferred example of Wasserstein distance on a persistent graph is as follows: For two persistent graphs D1 and D2, Where, the p-Wasserstein distance is: ; Where Δ is the diagonal set used to handle matching with unequal cardinality of point sets, the p-Wasserstein distance calculation method is used to eliminate the computational ambiguity of topological drift, D1 and D2 are the current persistent graph and the historical persistent graph, respectively, which can be used as a point set on the plane with diagonal set Δ={(a,a):a∈R}, and γ is a bijective matching from D1∪Δ to D2∪Δ. Chebyshev distance The norm, with parameter p taking either 1 or 2. The introduction of the diagonal set Δ ensures that the optimal matching cost between two persistent graphs with different cardinalities is always finite.

[0099] When a weak anomaly caused by an unknown eavesdropping signal appears in the electromagnetic environment, even if the electromagnetic radiation energy generated by the anomaly signal is extremely low, it will manifest as one or a few isolated points deviating from the normal point cloud on the positive definite matrix manifold. After continuous cohomology, the current persistent graph will show new transient generator points in the region far from the normal point cloud clustering area.

[0100] The electromagnetic detection module continuously calculates the Wasserstein distance of the current persistent graph at preset time intervals to obtain a topological drift sequence that varies over time. The topological drift sequence fluctuates slightly around a stable baseline under normal electromagnetic conditions, and the fluctuation amplitude reflects normal topological perturbations caused by legitimate random events in the background environment, such as Wi-Fi bursts and Bluetooth frequency hopping.

[0101] The electromagnetic detection module uses the Bayesian online change point detection method to monitor the topological drift sequence in real time. The Bayesian online change point detection establishes a probabilistic generation model for the drift sequence, assuming that the data follows different probability distributions before and after an unknown change point time, and uses Bayesian inference to update the posterior probability of the change point at the current time in real time. Whenever a new topological drift data point is acquired, the detector updates the posterior probability estimate once.

[0102] This application uses a Bayesian online change point detection method to monitor the topological drift sequence in real time. It does not require a preset fixed anomaly judgment threshold. It adaptively judges whether the current data is consistent with the past pattern based on the historical statistical characteristics of the sequence itself. When an unknown eavesdropping signal appears in the electromagnetic environment, the manifold point cloud disturbance caused by the unknown eavesdropping signal will cause a sudden increase in the mean or a change in the variance of the topological drift sequence. The Bayesian change point detector senses that the sequence statistical characteristics deviate from the historical operating pattern, and the posterior probability of the change point at the corresponding moment begins to rise.

[0103] The electromagnetic detection module has a preset threshold. When the posterior probability of a change point exceeds this threshold, it determines that a topological anomaly has occurred in the electromagnetic environment that cannot be explained by normal background fluctuations, and attributes it to the intervention of an unknown eavesdropping signal. At this time, the electromagnetic detection module outputs the posterior probability value of the change point at the current moment as a second threat indicator. The second threat indicator gives the degree of certainty that an anomaly exists in the current electromagnetic environment in probabilistic form.

[0104] When the posterior probability exceeds a preset confidence threshold, it is determined that a topological anomaly caused by an unknown eavesdropping signal has occurred in the electromagnetic environment, and the current posterior probability is output as the second threat indicator. A preferred example includes setting the confidence threshold to 0.85. like Figure 4 As shown, the fusion analysis module combines the first threat indicator and the second threat indicator with evidence under time alignment, and generates comprehensive alarm information according to preset spatiotemporal rules.

[0105] The first threat indicator and the second threat indicator are respectively converted into probability assignment functions under the same identification framework; By employing the evidence combination rule, the first basic probability assignment and the second basic probability assignment that are aligned in time are fused to obtain the joint confidence level allocation; The joint confidence level allocation is adjusted according to preset spatiotemporal logic rules, which include: when the laser eavesdropping event represented by the first threat indicator and the electromagnetic anomaly event represented by the second threat indicator overlap within a time window, the confidence level of the eavesdropping attack is increased; when only the second threat indicator appears and the first threat indicator does not simultaneously indicate laser eavesdropping, the confidence level of the eavesdropping attack is decreased and marked as an event to be observed, and comprehensive alarm information is generated and output according to the adjusted confidence level. In this application, a preferred implementation method of the fusion analysis module includes: The fusion analysis module predefines a unified identification framework, which contains several mutually exclusive propositions to describe the current security status of the classified space. For example, the identification framework may include three propositions: the existence of a data theft attack, the suspicious status requiring observation, and the normal environment.

[0106] The fusion analysis module converts the first threat indicator and the second threat indicator into basic probability assignment functions on the identification framework.

[0107] The first threat indicator is a scalar value used to reflect the degree of certainty that the vibration source module has about laser eavesdropping based on a combination of factors such as causal direction, time delay matching, and information flow intensity.

[0108] The fusion analysis module converts the value of the first threat indicator into a basic probability assignment for each proposition in the identification framework through a preset mapping relationship. The establishment of the mapping relationship is based on the detection statistical characteristics accumulated by the vibration source module during the testing and calibration phase: when the value of the first threat indicator is high, the basic probability assignment for the proposition of potential data theft increases, while the basic probability assignment for the proposition of normal environment decreases; when the value of the first threat indicator is below a certain confidence interval, the basic probability assignment is more allocated to the proposition of suspicious status requiring observation, so as to reflect the uncertainty of the detection results.

[0109] The second threat indicator is the posterior probability of the change point output by the electromagnetic detection module, which reflects the degree of certainty that the current electromagnetic environment topology is abnormal. The fusion judgment module maps the posterior probability value to the second basic probability assignment on the recognition framework. The two basic probability assignments independently express the degree of support of their respective detection channels for the current security status, and both allow a certain degree of confidence allocation to the uncertain parts of the overall recognition framework.

[0110] Furthermore, since the vibration source module and the electromagnetic detection module output indicators according to different processing cycles, there are sampling time differences between the two threat indicators on the time axis. The fusion analysis module first performs time alignment processing on the first threat indicator sequence and the second threat indicator sequence. For each time to be analyzed, the first threat indicator value and the second threat indicator value that are closest to the timestamp of that time are selected, or the aligned indicator pair is obtained by interpolation and time window aggregation. The aligned indicator pair corresponds to the same analysis time window.

[0111] After obtaining two sets of basic probability assignments that are time-aligned, the fusion analysis module uses evidence combination rules to fuse them. The evidence combination rules orthogonally synthesize the basic probability assignments of multiple independent evidence sources and calculate the joint confidence level allocation.

[0112] Define the recognition framework P1 and P2 represent the presence of a data theft attack, a suspicious state requiring observation, and a normal environment, respectively. P1 and P2 are mapped to basic probability assignments. m 1 and m 2 P1 is the primary threat indicator, and P2 is the secondary threat indicator.

[0113] A preferred example of orthogonal computation includes: ; ; Where K is the conflict factor. m 1 , m 2Let A, B, and C be the basic probability assignment functions for two independent sources of evidence, and let A, B, and C be propositions in the recognition framework.

[0114] After the evidence fusion is completed and the joint confidence level is assigned, adjustments are made according to the preset spatiotemporal logic rules.

[0115] If two threat indicators overlap within a time window, the risk level will increase. The confidence level; if only P2 occurs and P1 does not trigger an alarm, Confidence shift to When A comprehensive alarm is output when the confidence level is higher than the alarm threshold. A preferred alarm threshold example is 0.9.

[0116] During the synthesis process, the portion of two pieces of evidence that consistently support a proposition leads to an increase in the joint confidence of the proposition; the portions of two pieces of evidence that support contradictory propositions are transformed into conflict confidence, and are redistributed according to the selected conflict resolution strategy. After evidence combination, the fusion analysis module obtains a comprehensive joint confidence allocation, which integrates all evidence information from the laser eavesdropping detection channel and the electromagnetic anomaly detection channel.

[0117] The fusion analysis module further adjusts the joint confidence allocation based on preset spatiotemporal logic rules to achieve cross-channel collaborative verification and false alarm suppression. The spatiotemporal logic rules are formulated based on the temporal correlation characteristics of laser eavesdropping attacks and electromagnetic eavesdropping attacks in classified spaces.

[0118] When the laser eavesdropping event represented by the first threat indicator overlaps with the electromagnetic anomaly event represented by the second threat indicator within the time window, the confidence level of the espionage attack is increased. In the case of a composite espionage attack, the attacker uses both laser eavesdropping and electromagnetic espionage methods at the same time, or when the electromagnetic espionage device is triggered, there is a weak optical disturbance that can be perceived by the laser detection channel. Therefore, when the abnormal indicators of the two channels appear simultaneously in time, the possibility of a real attack increases. The enhancement operation specifically involves increasing the confidence level of propositions with potential data theft attacks in the joint confidence level allocation by a predetermined margin, while correspondingly reducing the confidence level of suspicious or normal propositions.

[0119] When only the second threat indicator appears while the first threat indicator does not simultaneously indicate laser eavesdropping, the confidence level of the eavesdropping attack is reduced and marked as an event to be observed. Topological anomalies in the electromagnetic environment may be caused by non-eavesdropping factors such as accidental external interference, legitimate new equipment access, or brief industrial electromagnetic pulses, lacking causal confirmation from the laser eavesdropping channel. Such isolated anomalies have low reliability. The adjustment method is to transfer most of the confidence of the proposition containing the eavesdropping attack in the joint confidence allocation to the proposition with a suspicious state requiring observation, avoiding high-frequency false alarms based on evidence from a single channel.

[0120] After the fusion analysis module completes the confidence adjustment according to the above rules, it takes the joint confidence of the proposition of data theft attack in the adjusted identification framework as the comprehensive alarm confidence level. The confidence level is compared with the preset alarm output threshold. When the confidence level is higher than the threshold, a comprehensive alarm information is generated and output, which may also include information such as alarm level, source of evidence of the channel on which it is based, and timestamp. When the confidence level is lower than the threshold, the current state is recorded but no alarm is triggered. Events marked as to be observed can be included in long-term statistics to assist in security situation assessment.

[0121] Example 2 like Figure 5 As shown, a method for sensing and analyzing multimodal data in classified spaces includes: Simultaneously acquire the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data within the classified space; Time-frequency transformations are performed on the first vibration signal, the second vibration signal, and the acoustic reference signal to obtain time-frequency spectra. Each time-frequency spectra is mapped to a regenerating kernel Hilbert space using a nonlinear kernel function. In the regenerating kernel Hilbert space, higher-order cross-cumulants of the three signals are calculated and a coupling tensor is constructed. The coupling tensor is subjected to regularization constraints and block decomposition to separate the time-domain components corresponding to multiple independent excitation sources; wherein the regularization constraints are derived from a fractional-order model characterizing the damping properties of the target glass, the fractional-order model includes fractional-order time derivative terms to characterize the viscoelastic memory effect of the glass, so that each separated excitation source physically conforms to the vibration transmission characteristics of the glass. Based on nonlinear transfer entropy and symbol transfer entropy, the directional information transmission between each separated excitation source and the first vibration signal is calculated. Excitation sources that simultaneously satisfy the causal direction of the first vibration signal pointing to the excitation source and the information transmission delay matching the preset laser eavesdropping delay range are marked as real laser eavesdropping sources. A first threat index is generated based on the directional information flow intensity and delay stability. A spatial covariance matrix is ​​constructed from the broadband electromagnetic spectrum data in time-frequency blocks to obtain multiple positive definite matrices and embed them into a positive definite matrix manifold equipped with affine invariant Riemannian metric. A filtered simple complex is constructed on the manifold based on geodesic distance and the persistent homology is calculated to generate the current persistent graph. Calculate the Wasserstein distance between the current persistent graph and the set of historical persistent graphs to obtain the topological drift sequence. Perform Bayesian online change point detection on the topological drift sequence. When the posterior probability of the detected change point exceeds a preset threshold, output the posterior probability as a second threat indicator. The first threat indicator and the second threat indicator are combined in time alignment, and the joint confidence is adjusted according to preset spatiotemporal logic rules to generate comprehensive alarm information.

[0122] Example 3 This embodiment provides a specific application example of a multimodal data perception and analysis system for classified spaces. Taking a classified conference room as the protected object, it continuously monitors the window glass area for laser eavesdropping and electromagnetic espionage threats.

[0123] In this embodiment, the confidential conference room has an exterior glass window with a double-layered laminated glass structure and a thickness of 6mm. Offline impact testing revealed its fractional-order model parameters as follows: density ρ is taken as 2500 kg / m³ converted to mass per unit area, damping coefficient c is 120 N·s / m, and stiffness parameter k is 8.5 × 10⁻⁶. 4 N / m, fractional order α is taken as 0.48.

[0124] The first vibration signal was acquired by a laser Doppler vibrometer. The laser spot of the vibrometer was pointed towards the center of the outer surface of the glass window, with a spot diameter of approximately 0.5 mm. The vibrometer output a vibration velocity signal with a sampling rate set to 100 kHz. The second vibration signal was acquired by a piezoelectric accelerometer. The accelerometer was attached to the inner surface of the same glass surface via a coupling agent, and its installation position was coaxial with the vibration spot. The accelerometer sensitivity was 100 mV / g, and the output was synchronously acquired at the same 100 kHz sampling rate after being amplified by a charge amplifier. The acoustic reference signal was acquired by an indoor microphone installed on the conference table 1.5 m above the glass window at a height of 1.2 m, picking up indoor speech and ambient sound fields. Broadband electromagnetic spectrum data was acquired by four broadband receiving antennas covering the 1 MHz to 6 GHz frequency band. The four antennas were arranged in a rectangular array at the four corners of the conference room. Each channel was synchronously down-converted and analog-to-digital converted, outputting a complex baseband data stream with a sampling rate of 50 MSPS.

[0125] The vibration source module performs short-time Fourier transform on the first vibration signal, the second vibration signal, and the acoustic reference signal with a window length of 1024 points and an overlap of 512 points to obtain three time-frequency spectra. The Gaussian radial basis kernel function is used to map each time-frequency spectrum point to the regenerating kernel Hilbert space. The kernel parameters are selected through cross-validation. The third-order joint cumulant of the three signals is calculated in the regenerating kernel Hilbert space to construct a coupling tensor. The three dimensions of the tensor correspond to the time-frequency grid indexes of the first vibration signal, the second vibration signal, and the acoustic reference signal, respectively.

[0126] In the block decomposition stage, the vibration source module represents the coupling tensor as the sum of several low-rank tensor blocks. The number of excitation sources R is selected based on the multilinear rank distribution of the coupling tensor and the expected number of potential excitation sources in the actual scenario. In this embodiment, R is set to 3, corresponding to indoor voice sources, laser eavesdropping related sources, and environmental pseudo-vibration sources, respectively. A physical regularization term is introduced into the decomposition objective function. The fractional derivative in the regularization term is calculated using a discrete approximation. The discrete truncation length N is set to 200 points, and the tradeoff parameter λ is set to 0.1. The solution is obtained iteratively using the alternating direction multiplier method, with a maximum number of iterations set to 500 and a convergence threshold set to 10. -6 Finally, the projection time-domain waveforms of the three independent excitation sources on the first vibration signal, the second vibration signal, and the acoustic reference signal are obtained.

[0127] The vibration source module calculates the nonlinear transfer entropy between itself and the first vibration signal. The embedding dimensions m and n are determined to be 3 and 4 respectively using the spurious nearest neighbor method. The embedding delay Δt is taken as 5 sampling points. The symbolic transfer entropy based on the sorting entropy is calculated synchronously. The sorting pattern length is 4. The statistical significance of the transfer entropy is determined by 1000 random permutation tests, and the significance level is 0.01. For excitation sources that simultaneously satisfy the causal direction of the first vibration signal pointing to the excitation source and the estimated time delay is in the range of 0.02ms to 0.5ms, they are marked as real laser eavesdropping sources. In this embodiment, the preset delay range of the laser eavesdropping process is determined by combining the optical path round-trip distance of about 3m and the vibration propagation time inside the glass. The transfer entropy value of the marked source is obtained as the directional information flow intensity, and the standard deviation of the time delay estimate is calculated as a stability measure within 20 consecutive time windows. The two are weighted and summed to generate the first threat index. The weighting coefficients are determined to be 0.7 and 0.3 through previous calibration experiments.

[0128] During operation, the electromagnetic detection module divides the broadband electromagnetic spectrum data into time frames of 5ms each and frequency sub-bands of 100kHz each. Each time-frequency block contains complex baseband sampling data from four antennas, and a 4×4 sample covariance matrix is ​​calculated. Based on the actual rectangular geometry of the antenna array, the sample covariance matrix is ​​corrected using Toeplitz structural constraints, and a regularization value of 10 is added to the diagonal elements. -8 To ensure the positive definiteness of the matrix, the spatial covariance matrix of each time-frequency block is obtained.

[0129] The electromagnetic detection module assigns an affine-invariant Riemannian metric to each spatial covariance matrix. A filtered simplex is constructed on a positive definite matrix manifold using geodesic distances. The distance threshold starts at 0.1 times the minimum geodesic distance and increases proportionally to 5 times the maximum geodesic distance, with a total of 50 threshold levels. The module calculates the continuous homology of the filtered sequence, records the birth and death parameters of 0-dimensional and 1-dimensional topological features, and generates a persistent graph for the current monitoring time window. The historical background persistent graph set is composed of persistent graphs acquired during the initial 10 consecutive working days after system deployment, and is maintained and updated using a 30-day sliding window.

[0130] The electromagnetic detection module calculates the 1-Wasserstein distance between the current persistent graph and the historical background persistent graph set every 10 seconds, forming a topological drift sequence. A Bayesian online change point detection algorithm is applied to this sequence, assuming that the probability distribution of the drift before and after the change point is Gaussian. Prior parameters are estimated using historical normal data. Each time a new topological drift data point is acquired, the posterior probability of the current time being a change point is updated. In this embodiment, the confidence threshold is set to 0.85. When the posterior probability exceeds 0.85, a topological anomaly is determined to exist in the electromagnetic environment, and the current posterior probability value is output as the second threat indicator.

[0131] The fusion analysis module defines the identification framework as {espionage attack, suspicious state, normal environment}. It converts the first threat indicator into a first basic probability assignment using a preset piecewise linear mapping function: when the first threat indicator value is above 0.8, the basic probability assignment for the eavesdropping attack proposition is 0.80, for the suspicious state it is 0.15, and for the normal environment it is 0.05; when the value is below 0.3, the assignment is 0.10 for the eavesdropping attack proposition, 0.60 for the suspicious state it is, and 0.30 for the normal environment. The second threat indicator, i.e., the posterior probability of the change point, is directly used as the degree of support for the electromagnetic anomaly. The second basic probability assignment assigns the posterior probability value to the eavesdropping attack proposition, and the remaining confidence is assigned to the suspicious state proposition.

[0132] The fusion analysis module aligns the first and second threat indicators within a 10-second analysis window, performs orthogonal synthesis using evidence combination rules, and handles the conflict factor K using a standard normalization method. After fusion, adjustments are made according to preset spatiotemporal logic rules: when the laser eavesdropping event represented by the first threat indicator and the electromagnetic anomaly represented by the second threat indicator coexist within a 5-second overlap window, the confidence of the eavesdropping attack proposition in the joint confidence is multiplied by a coefficient of 1.3, and correspondingly deducted from the state-suspect proposition; when only the second threat indicator appears and the first threat indicator does not alarm synchronously, 80% of the confidence of the eavesdropping attack proposition is transferred to the state-suspect proposition and marked as an event to be observed. After adjustment, when the joint confidence level of the data theft attack proposition is higher than the alarm threshold of 0.90, the system outputs a first-level comprehensive alarm message, with an alarm timestamp, contributing channel identifier and confidence level; when the confidence level is between 0.70 and 0.90, a second-level warning message is output and continuously tracked; when it is lower than 0.70, only the current status is recorded to the security log and no alarm is triggered.

[0133] In this embodiment, the present application successfully detected laser eavesdropping sources and electromagnetic anomalies in a simulated composite attack scenario, and increased the confidence level of the eavesdropping attack to 0.94 through evidence fusion, outputting a first-level comprehensive alarm. In another scenario where only temporary wireless devices were connected, the second threat indicator briefly increased, but since the first threat indicator did not trigger an alarm, the spatiotemporal logic rules adjusted the confidence level to 0.45 and marked it as pending observation, without triggering a false alarm, thus verifying the effectiveness and reliability of the system in practical applications.

[0134] It is important to note that the constructions and arrangements of this application shown in several different exemplary embodiments are merely illustrative. Although only two embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible without substantially departing from the novel teachings and advantages of the subject matter described in this application. These modifications may include, for example, changes in the size, dimensions, structure, shape, and proportions of various elements, as well as parameter values ​​(e.g., temperature, pressure, etc.), installation arrangements, the use of materials, colors, orientations, etc. For example, an element shown as integrally formed may be composed of multiple parts or elements, the position of elements may be inverted or otherwise altered, and the nature or number or position of discrete elements may be changed or altered. Therefore, all such modifications are intended to be included within the scope of this application. The order or sequence of any process or method may be changed or reordered by alternative embodiments. Any "apparatus plus function" clause is intended to cover, and not only structurally equivalent but also equivalent structures, the structures performing the functions described herein. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of this application. Therefore, this application is not limited to a particular embodiment, but extends to various modifications that still fall within the scope of the appended claims.

[0135] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of performing this application as currently considered, or those features that are not relevant to implementing this application) may be omitted.

[0136] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those of ordinary skill in the art who benefit from this disclosure, the development effort will be a routine task in design, manufacturing, and production without requiring extensive experimentation.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application, and all such modifications and substitutions should be covered within the scope of the claims of this application.

Claims

1. A system for sensing and analyzing multimodal data in classified spaces, characterized in that, include: The multimodal module is used to simultaneously acquire the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data; The vibration source module maps the time-frequency representation of the first vibration signal, the second vibration signal, and the acoustic reference signal, constructs a coupling tensor, applies a regularization constraint to the coupling tensor, and performs block term decomposition to separate the time-domain components of multiple excitation sources. Through the transmission metric between the excitation source and the first vibration signal, the laser eavesdropping source is identified and the first threat indicator is output. The electromagnetic detection module constructs a spatial covariance matrix based on the time-frequency blocks of the broadband electromagnetic spectrum data to generate the current persistent graph, and calculates the distance between the current persistent graph and the historical persistent graph as the topology drift. The change point detects the topology drift and outputs the second threat indicator. The fusion analysis module combines the first threat indicator and the second threat indicator with evidence aligned to time, and generates comprehensive alarm information according to preset spatiotemporal rules.

2. The system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: The first vibration signal was collected by a non-contact optical vibration measuring device pointing towards the window glass; The second vibration signal was acquired by a contact vibration sensor attached to the same glass. The acoustic reference signal is acquired by an indoor microphone; The broadband electromagnetic spectrum data is collected by multiple broadband receiving antennas.

3. The system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: Time-frequency transformations were performed on the first vibration signal, the second vibration signal, and the acoustic reference signal to obtain time-frequency spectra. Each time-frequency spectrum is mapped to the regenerating kernel Hilbert space by a nonlinear kernel function to obtain the high-order nonlinear similarity between each modal signal that cannot be linearly expressed in the original time-frequency domain. In the regenerative kernel Hilbert space, the higher-order cross-cumulants of the first vibration signal, the second vibration signal, and the acoustic reference signal are calculated, and a coupling tensor is constructed such that the elements of the coupling tensor represent the statistical correlation between different modes at different time and frequency points.

4. The system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: The method for applying regularization constraints and performing block decomposition to the vibration source module is as follows: The regularization constraint originates from a fractional-order model characterizing the damping properties of the target glass, which includes a fractional-order time derivative term to represent the viscoelastic memory of the classified space glass. Using the fractional-order model, a physical regularization term is constructed, which penalizes deviations from the fractional-order differential relationship that each separated excitation source should satisfy between the first vibration signal, the second vibration signal, and the acoustic reference signal. During the block decomposition process, the coupling tensor is represented as the sum of multiple low-rank tensor blocks, each tensor block corresponding to an independent excitation source, and the physical regularization term is introduced into the decomposition objective function so that the separation result is forced to follow the fractional order law of the glass. By solving the block decomposition with physical regularization constraints, multiple independent time-domain components of the excitation source that physically conform to the glass vibration transmission characteristics are obtained.

5. A system for sensing and analyzing multimodal data in classified spaces as described in claim 4, characterized in that: The method for identifying the laser eavesdropping source and outputting the first threat indicator based on the transmission metric between the excitation source and the first vibration signal is as follows: The transfer measure includes nonlinear transfer entropy and symbol transfer entropy based on sorting entropy; Calculate the information transmission between each separated excitation source and the first vibration signal to determine whether there is a causal drive from the first vibration signal to the excitation source, and estimate the time delay of information transmission; An excitation source that simultaneously satisfies the causal direction of the first vibration signal pointing to the excitation source and the time delay matching the preset delay range of the laser eavesdropping process is marked as a real laser eavesdropping source, and the directional information flow intensity and information transmission time delay of the real laser eavesdropping source are obtained. The first threat indicator is generated by measuring the strength of the directional information stream and the stability of its delay, which are identified as real laser eavesdropping sources.

6. A system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: The electromagnetic detection module is configured to divide the broadband electromagnetic spectrum data into several time-frequency blocks according to time frames and frequency subbands; calculate the sample covariance matrix of the multi-channel complex baseband sampling data in each time-frequency block; apply structural constraints to the sample covariance matrix and correct it to a positive definite matrix as the spatial covariance matrix of the time-frequency block; embed multiple spatial covariance matrices into a positive definite manifold, construct a filtered simplex complex on the positive definite manifold based on geodesic distance and calculate the persistent homology to generate the current persistent graph.

7. A system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: The electromagnetic detection module is further configured to: assign an affine invariant Riemannian metric to each of the spatial covariance matrices, such that all the positive definite matrices to be processed form a positive definite matrix manifold with an affine invariant Riemannian metric; Using geodesic distance as the distance metric between two points on the positive definite matrix manifold, a simplex is constructed for filtering. As the distance threshold gradually increases, edges, triangles, and higher-order simplexes are generated sequentially between points on the manifold whose distance is less than the current threshold, forming a simplex filtering sequence that evolves with the threshold. Calculate the persistent homology of the filtering sequence, record the occurrence and extinction times of each topological feature during the filtering process, and generate a persistent graph labeled with birth and death parameter pairs. The persistent graph of the current monitoring time window is used as the current persistent graph, and each point of the current persistent graph represents a stable generator in the electromagnetic environment topology.

8. A system for sensing and analyzing multimodal data in classified spaces as described in claim 1, characterized in that: A collection of persistent graphs acquired multiple times under normal electromagnetic environments throughout history was constructed. Using the Wasserstein distance between persistent graphs as a metric, the topological difference between the current persistent graph and the set of historical persistent graphs is calculated to obtain a sequence of topological drift over time. Perform Bayesian online change point detection on the topological drift sequence and estimate the posterior probability of the current time being an anomalous change point in real time; When the posterior probability exceeds a preset confidence threshold, it is determined that a topological anomaly caused by an unknown eavesdropping signal has occurred in the electromagnetic environment, and the current posterior probability is output as the second threat indicator.

9. A system for sensing and analyzing multimodal data in classified spaces as described in claim 8, characterized in that: The first threat indicator and the second threat indicator are respectively converted into probability assignment functions under the same identification framework; By employing the evidence combination rule, the first basic probability assignment and the second basic probability assignment that are aligned in time are fused to obtain the joint confidence level allocation; The joint confidence level allocation is adjusted according to a preset spatiotemporal logic rule, which includes: when the laser eavesdropping event represented by the first threat indicator and the electromagnetic anomaly event represented by the second threat indicator overlap within a time window, the confidence level of the eavesdropping attack is increased; when only the second threat indicator appears and the first threat indicator does not simultaneously indicate laser eavesdropping, the confidence level of the eavesdropping attack is reduced and marked as an event to be observed.

10. A method for sensing and analyzing multimodal data in classified spaces, characterized in that, A system for sensing and analyzing multimodal data in classified spaces as described in any one of claims 1 to 9, comprising: S1 simultaneously acquires the first vibration signal, the second vibration signal, the acoustic reference signal, and broadband electromagnetic spectrum data; S2 processes the first vibration signal, the second vibration signal, and the acoustic reference signal to construct a higher-order coupling tensor. It applies physical regularization constraints to the coupling tensor and performs block term decomposition to separate the time-domain components of multiple excitation sources. Based on the nonlinear directional information transmission measure between each excitation source and the first vibration signal, it identifies laser eavesdropping sources and generates a first threat index. S3 processes the broadband electromagnetic spectrum data to construct a spatial covariance matrix and embed it into a positive definite manifold. A filtered complex is constructed on the manifold, and a continuous homology is calculated to generate the current persistent graph. The topological drift between the current persistent graph and the historical persistent graph is calculated, and a second threat index is generated through change point detection. S4 performs time-aligned evidence fusion on the first threat indicator and the second threat indicator to generate comprehensive alarm information.

Images