Public security AI-driven data generation type investigation teaching training system

Abstract

The invention relates to the field of training simulation, and discloses a public security AI-driven data generation type investigation teaching practical training system, which comprises the following modules: a case dynamic generation module, which is used for receiving original data of a public security case database, generating a logically closed virtual case through tensor decomposition, and sending the virtual case to the public security case database; transmitting the generated case data to a multi-modal virtual-real fusion training module; and the multi-modal virtual-real fusion training module is used for receiving the virtual case data, realizing interactive training of multi-modal clues through a mixed reality technology, dynamically injecting interference clues, and transmitting student operation data to the causal reinforcement learning evaluation module. Through the collaborative technical scheme of tensor decomposition and nonlinear constraint optimization, the integrity and space-time consistency generation of the virtual case evidence chain is realized, and compared with a traditional case generation method based on a fixed template, the problem of contradiction between material evidence association fracture and behavior logic is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of training simulation technology, and in particular to a public security AI-driven data-generated investigation teaching and training system. Background Art

[0002] Currently, telecom fraud and dark web trading crimes have generally adopted AI face-changing and blockchain anonymous transfer technologies, while the traditional case database is still dominated by traditional theft and robbery cases. When police officers face deep fake suspect voiceprints and new evidence of cryptocurrency fund flows in real cases, the lack of targeted training has led to an increase in the rate of missed detection of key clues.

[0003] Traditional case generation relies on manually preset templates. Template solidification limits the dynamic association of case elements, and physical evidence and suspect behavior often conflict in time and space. When the surveillance video shows that suspect A was in place B at the time of the crime, but in the generated case, A holds physical evidence in place C. Trainees are trained in contradictory scenarios, which can easily lead to incorrect investigation path dependence. Traditional assessments focus on the completeness of operational steps and ignore the causal relationship between behavior and results. When trainees complete physical evidence collection according to a fixed process but fail to trigger the conditions for solving the case, it is difficult for the system to identify the missing key operations. This lack of feedback leads to repeated incorrect operations. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a public security AI-driven data-generation investigation teaching and training system, which solves the problems in the existing technology that static encryption and access control strategies cannot flexibly adapt to dynamic data protection needs, waste resources and lag in compliance.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a data generation-based investigation teaching and training system driven by public security AI, comprising the following modules:

[0006] The case dynamic generation module is used to receive the original data of the public security case database, generate logically closed virtual cases through tensor decomposition, and transmit the generated case data to the multi-modal virtual-real fusion training module;

[0007] A multimodal virtual-reality fusion training module is used to receive the virtual case data, implement interactive training of multimodal clues through mixed reality technology, dynamically inject interference clues, and transmit the trainee operation data to the causal reinforcement learning evaluation module;

[0008] A causal reinforcement learning evaluation module is used to receive the student's operation data, construct a time-series causal graph and quantify the causal contribution, generate an evaluation report and feed it back to the case dynamic generation module to adjust the case generation parameters;

[0009] A fully homomorphic encryption security module is used to encrypt and decrypt the transmission data in real time between the case dynamic generation module, the multi-modal virtual-real fusion training module, and the causal reinforcement learning evaluation module;

[0010] The public security business extension interface module receives real-time case feature data from the public security actual combat platform and imports it into the case dynamic generation module, and at the same time exports the generated case solutions to the police command system

[0011] Preferably, the case dynamic generation module includes:

[0012] The case element preprocessing unit accesses the public security case database, obtains structured data (suspect files, physical evidence lists) and unstructured data (interrogation transcripts, surveillance videos, communication records), extracts case elements (suspects, time, location, physical evidence) through natural language processing (NLP), constructs a knowledge graph, where nodes represent entities and edges represent association relationships ("Suspect A contacted physical evidence B at time T"), divides the place where the case occurred into grids (100m×100m) according to longitude and latitude, and slices the time by hour to form four-dimensional spatio-temporal coordinates;

[0013] The case element tensor encoding unit defines the tensor dimension as (suspect × physical evidence × time × space), the tensor element value is the association strength (fingerprint matching degree, DNA similarity) between the suspect and the physical evidence in this spatio-temporal context, decomposes the tensor into a core tensor and factor matrices, extracts potential association patterns (certain types of physical evidence appear frequently in a specific time period), sets non-linear constraint conditions (the physical evidence timeline is continuous, the suspect's trajectory is reasonable), and generates a virtual case with a closed evidence chain through the gradient descent method;

[0014] The non-linear constraint optimization unit constructs and solves the investigation logic constraint equations based on the decomposition results to generate a virtual case with a closed evidence chain;

[0015] The adversarial behavior simulation unit trains an AI suspect agent to simulate anti-investigation behaviors such as forging physical evidence (generating false surveillance timestamps) and creating contradictory confessions (using GPT-2 to generate confusing texts), and adjusts the suspect's behavior strategy in real time according to the trainee's operations (increasing the frequency of physical evidence forgery) to generate multi-branch case deduction paths.

[0016] Tensor decomposition mines the potential associations between case elements (the relationship between the suspect's behavior pattern and the distribution of physical evidence) through dimensionality reduction, breaking through the static limitations of traditional rule bases. Logic constraint optimization ensures the self-consistency of the evidence chain of the virtual case (avoiding the same physical evidence appearing in contradictory spatio-temporal contexts), and adversarial behavior simulation enhances the complexity of the case through dynamic strategies, making the training scenario closer to the real crime evolution law. The virtual cases generated by this module are transmitted to the multi-modal virtual-real fusion training module through a standardized interface to support the construction of an interactive training scenario.

[0017] Preferably, the multi-modal virtual-real fusion training module includes:

[0018] A cross-modal alignment engine that uses a pre-trained model (CLIP) to map the transcript text and surveillance images to the same feature space, calculates the semantic similarity to match associated clues, and verifies the logical consistency of physical evidence locations, timestamps, and video segments through a graph neural network (GNN) (the time when the suspect appears in the surveillance footage matches the mobile phone location);

[0019] A dynamic variable injector that superimposes spatio-temporal environmental variables during the diffusion model generation process, uses the Unity engine to construct a three-dimensional virtual scene, superimposes real environmental data (weather, lighting), and allows trainees to conduct immersive investigations through AR glasses. During the training process, environmental interferences are randomly inserted (simulating rainy days affecting fingerprint collection and reduced visibility at night) to enhance the authenticity and challenge of the scene;

[0020] An interference clue generator that mixes in interference clues (forged physical evidence labels, contradictory witness testimonies) at a preset ratio (30%), simulates the investigative interference created by the suspect, integrates text (call records), images (scene photos), and time-series data (fund flow) into a unified training task, and requires trainees to conduct cross-modal correlation analysis.

[0021] Cross-modal alignment ensures the logical consistency of clues through semantic mapping (the association between text descriptions and image content), avoiding distortion of the training scene. The dynamic variable injection simulates environmental interferences in real investigations (the impact of bad weather on evidence collection), forcing trainees to adapt to complex conditions. The game strategy of interference clues trains trainees' information screening and logical reasoning abilities by confusing true and false information. The operation data of trainees collected by this module (the order of clue retrieval, decision-making time) is transmitted to the causal reinforcement learning evaluation module in real time to form a training-evaluation closed loop.

[0022] Preferably, the causal reinforcement learning evaluation module includes:

[0023] A temporal causal graph construction unit that extracts the causal dependencies of trainees' operation sequences through a temporal convolutional network, encodes trainees' actions (retrieving surveillance footage, interrogating suspects, applying for search warrants) as time-series vectors, extracts long-term and short-term dependencies through a temporal convolutional network (TCN), identifies key causal chains ("DNA comparison successful → suspect locked → case solved"), and eliminates invalid operations (repeated retrieval of the same physical evidence);

[0024] A causal contribution quantification unit that calculates the causal effect of an operation on the case-solving result, calculates the result difference when a certain operation is not executed (failure to freeze the account leads to fund transfer), quantifies the causal effect value of the operation node (in the range of 0-1), and outputs the operation effectiveness score (the accuracy rate of key physical evidence identification is 90%), error points (omission of cross-provincial IP association), and optimization suggestions (shortening the response time).

[0025] The meta - learning dynamic adaptation unit dynamically adjusts the difficulty of case generation based on the evaluation results. Based on historical evaluation data (average scores of trainees, common types of mistakes), it adaptively adjusts case parameters (increasing the technical complexity of physical evidence forgery techniques), reduces the proportion of interfering clues in the novice stage, and simulates complex cases such as gang crimes and cross - border money laundering in the advanced stage.

[0026] Sequential causal modeling breaks through the limitations of traditional statistical evaluation by capturing the dependency relationships in the operation chain (the failure of subsequent processes due to the absence of a certain step). The causal contribution quantification model uses counterfactual reasoning (the difference in results assuming a certain operation is not executed) to accurately locate key decision points. Dynamic difficulty adaptation realizes personalized training path planning through a meta - learning algorithm (adjusting case parameters based on the trainee's ability profile). The evaluation results are transmitted back to the case dynamic generation module in real - time through the feedback link, forming a "training - evaluation - optimization" closed - loop.

[0027] Preferably, the fully homomorphic encryption security module includes:

[0028] The real - time encryption and decryption unit encrypts case data and trainee operation records using the BFV scheme, supports data transmission and calculation in the ciphertext state, generates a session key based on the timestamp hash chain (updated every 5 minutes), and stores the main key in combination with a hardware security module (HSM) to prevent key leakage.

[0029] The security audit and compliance unit records data visitors, operation types, timestamps, and encrypts and stores them in the blockchain (Hyperledger Fabric) to ensure that the logs cannot be tampered with, sets three - level permissions for trainees, instructors, and administrators, and restricts the access scope of sensitive data (real case information).

[0030] Fully homomorphic encryption allows operations to be performed in the ciphertext state (correlation analysis of encrypted clues), avoiding the vulnerability risk of traditional encryption technologies that require frequent decryption. The dynamic key mechanism resists replay attacks through timeliness control (periodically updating the session key), and the hardware security module (HSM) ensures the physical isolation of the main key. The blockchain audit mechanism ensures the integrity and traceability of operation logs, meeting the compliance requirements of the "Public Security Organs Data Security Specification". This module serves as the system security foundation, ensuring data confidentiality throughout the entire process of case generation, training, and evaluation.

[0031] Preferably, the public security business extension interface module includes:

[0032] The actual combat data import unit accesses new - type crime data (virtual currency money - laundering models, dark web trading characteristics) from the police command platform, updates the case generation template, and feeds back the investigation strategies (fund flow tracking algorithms) that have been verified effective in actual combat to the system to optimize the virtual case generation logic.

[0033] The training results output unit converts the tactical processes verified in the training (electronic evidence collection steps, suspect tracking paths) into standard operation manuals and pushes them to the police APP. Before solving major cases, the system simulates the effects of various investigation plans (encirclement routes Avs. B) to assist the command center in making decisions.

[0034] The protocol compatibility design unit supports the GA / T1400 video transmission protocol and the GB / T28181 networking protocol, realizes data interoperability with the Skynet Project and the Xueliang Project, provides RESTful API and SDK, and supports rapid connection with third-party systems (mobile alarm platform, intelligence analysis system).

[0035] Real-time case feature import ensures that virtual cases are synchronized with new crime trends (simulating the latest fraud tactics), and the reverse output of training results (tactical manual) directly empowers front-line police work and solves the problem of the disconnection between traditional training and actual combat. The protocol compatibility design realizes seamless cross-system data flow through standardized interfaces (GA / T1400) to avoid data silos. As a bridge for the integration of combat and training, this module promotes the continuous iteration and upgrading of investigation tactics.

[0036] Preferably, the case element tensor encoding unit performs Tucker decomposition to satisfy;

[0037]

[0038] in:

[0039] is the core tensor, representing the potential correlation across dimensions;

[0040] are the factor matrices of suspect attributes and physical evidence types respectively;

[0041] are the time and space factor matrices respectively.

[0042] Preferably, the system of equations solved by the nonlinear constraint optimization unit is:

[0043]

[0044] in:

[0045] x i ∈[0,1]: the activation probability of the i-th element in the case element vector;

[0046] α i ∈[0,1]: Evidence weight coefficient, calculated by the frequency of similar physical evidence in historical cases;

[0047] β = 0.8: The evidence chain closure threshold, indicating that at least 80% of the key evidence needs to be associated;

[0048] Spatio-temporal correlation matrix, element a ij Indicates the spatial reachability between positions i and j;

[0049] Benchmark vector, representing the standard spatio-temporal distribution pattern;

[0050] ∈ = 0.1: Tolerance parameter, allowing the spatio-temporal distribution deviation not to exceed 10%.

[0051] Preferably, the loss function of the cross-modal alignment engine is:

[0052]

[0053] Where:

[0054] λ = 0.8 is the balance coefficient,

[0055] p text = Softmax(φ text (s i ) / τ), p image = Softmax(Wφ image (v i ) / τ), and the temperature coefficient τ = 0.07.

[0056] Preferably, the dynamic variable injector executes in the diffusion model:

[0057] v t = LSTM(d1,…,d T ), d t = [temperature t , humidity t , light t

[0058] Where:

[0059] The dimension of the hidden layer of the LSTM network is 128, and the time step T = 24;

[0060] The noise scheduling parameter α of the diffusion model t Decays according to the cosine rule.

[0061] Preferably, the causal contribution quantification unit satisfies:

[0062]

[0063] Where:

[0064] ​Is an indicator function that outputs 1 when the operation exists and 0 otherwise;

[0065] Counterfactual sample Generated by GAN, and the generator is a 3-layer MLP.

[0066] Preferably, the fully homomorphic encryption security module performs the following operations:

[0067] Encrypted tensor operation: Perform homomorphic encryption on the tensor data in case generation and evaluation, satisfying:

[0068]

[0069] Where the CPD decomposition rank R = 32, the encryption scheme is CKKS, and the polynomial modulus degree N = 8192;

[0070] Edge-cloud collaboration: Edge nodes process real-time interactive data, and the cloud performs offline model training;

[0071] Key dynamic management: Rotate the encryption key every 24 hours and protect historical data through a forward security protocol.

[0072] The present invention provides a public security AI-driven data generation-based investigation teaching and training system. It has the following beneficial effects:

[0073] 1. Through the collaborative technical solution of tensor decomposition and non-linear constraint optimization, the present invention realizes the integrity and spatio-temporal consistency generation of the virtual case evidence chain. Compared with the traditional case generation method based on a fixed template, it effectively solves the problems of broken physical evidence association and behavioral logic contradiction, and ensures that the investigation and reasoning process in the training scenario conforms to the real case handling specifications.

[0074] 2. By adopting the cross-modal semantic alignment and dynamic environment variable injection technology, the present invention overcomes the semantic fragmentation problem of text and image clues in the virtual scene. Compared with a single-modal training system, it significantly improves the perceptual authenticity of clue association in the mixed reality environment, and specifically strengthens the comprehensive judgment ability of trainees in complex environments.

[0075] 3. Through the counterfactual reasoning mechanism based on causal reinforcement learning, the present invention constructs a dynamic quantitative association model between operation behaviors and case-solving results. Compared with the traditional static scoring system, it realizes the real-time matching of training difficulty and trainee ability, and avoids the common defects of overloading training and inefficient repetition.

[0076] 4. By innovatively integrating the fully homomorphic encryption tensor operation and noise dynamic management technology, while ensuring the privacy of investigation teaching data, the present invention breaks through the bottleneck of ciphertext operation efficiency. Compared with the conventional segmented encryption scheme, it takes into account both data security and system response speed in high-concurrency training scenarios. Brief Description of the Drawings

[0077] Figure 1 is the overall schematic diagram of the system of the present invention;

[0078] Figure 2 is the schematic diagram of the case dynamic generation module of the present invention;

[0079] Figure 3 is the schematic diagram of the multi-modal virtual-real fusion training module of the present invention;

[0080] Figure 4 is the schematic diagram of the causal reinforcement learning evaluation module of the present invention;

[0081] Figure 5 is the schematic diagram of the fully homomorphic encryption security module of the present invention;

[0082] Figure 6 is the schematic diagram of the public security business expansion interface module of the present invention. Detailed implementation manners

[0083] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0084] Please refer to the attached Figure 1-6 , the embodiments of the present invention provide a public security AI-driven data generation-based investigation teaching and training system, including a case dynamic generation module, which is used to receive the original data of the public security case database, generate a logically closed virtual case through tensor decomposition, and transmit the generated case data to the multi-modal virtual-real fusion training module. The case dynamic generation module includes a case element preprocessing unit, and the case element preprocessing unit includes a data cleaning part and an entity relationship extraction part. The data cleaning part receives the original unstructured data (text transcripts, surveillance videos, physical evidence records) of the public security case database, removes noise data through regular expression matching and outlier filtering, and retains structured fields including the time of the crime, geographical coordinates, basic attributes of the suspect, and types of physical evidence. The entity relationship extraction part uses a bidirectional long short-term memory network-conditional random field (BiLSTM-CRF) model to identify suspect entities (name, age, occupation) and physical evidence entities (fingerprints, DNA, electronic data). The input of the model is the text sequence after word segmentation, and the output is the entity label (BIO annotation system). Further, a triple relationship (head entity, relationship, tail entity) is constructed based on the TransE algorithm, and the relationship types include "appears in", "holds", and "is associated with". The embedding vector dimension is set to 128, and the loss function is:

[0085]

[0086] Where S is the set of valid triples, and h, r, and t are the embedding vectors of the head entity, relation, and tail entity respectively. The preprocessed structured data is output to the case element tensor encoding unit.

[0087] The case dynamic generation module includes a case element tensor encoding unit, which maps the preprocessed data into a fourth-order tensor Its dimensions are defined as follows:

[0088] Suspect attribute dimension I: Divided into age segments (18 - 25, 26 - 35, 36 - 45, 46 - 55, 56+), occupation types (unemployed, service industry, technicians, management, others), and behavior patterns (premeditated, impulsive, gang type), a total of 5×5×3 = 75 categories;

[0089] Physical evidence type dimension J: Divided into 8 categories including biological evidence (DNA, fingerprints), electronic data (communication records, transaction logs), and trace evidence (shoe prints, tool marks);

[0090] Time slice dimension K: Divided by the hour, and cases with a duration exceeding 72 hours are divided into multiple sub - cases;

[0091] Spatial grid dimension L: Using Geohash encoding (precision level 6, grid side length ≈ 1.2 km), covering the area within 5 kilometers around the crime scene.

[0092] Extract potential association patterns through Tucker decomposition, and the decomposition formula is:

[0093]

[0094] Where:

[0095] is the core tensor, representing the potential association across dimensions;

[0096] are the factor matrices of suspect attributes and physical evidence types respectively;

[0097] are the factor matrices of time and space respectively.

[0098] The decomposition process is iteratively optimized by the alternating least squares method (ALS), and the convergence condition is:

[0099]

[0100] The decomposition result is transmitted to the non - linear constraint optimization unit.

[0101] The optimization unit constructs the following constraint equations to ensure the logical closure and spatio-temporal consistency of case generation:

[0102]

[0103] Parameter definitions and calculation relationships:

[0104] x i ∈[0,1]: The activation probability of the i-th element in the case element vector;

[0105] α i ∈[0,1]: The evidence weight coefficient, calculated from the occurrence frequency of the same type of physical evidence in historical cases (DNA evidence weight α DNA =0.9, fingerprint evidence weight α 指纹 =0.7);

[0106] β = 0.8: The evidence chain closure threshold, indicating that at least 80% of the key evidence associations need to be satisfied;

[0107] Spatio-temporal correlation matrix, element a ij represents the spatial accessibility between positions i and j (calculated based on GIS data, accessibility = 1 - normalized path distance);

[0108] Benchmark vector, representing the standard spatio-temporal distribution pattern (case occurrence time concentration rate ≥ 70%);

[0109] ∈ = 0.1: Tolerance parameter, allowing the spatio-temporal distribution deviation not to exceed 10%.

[0110] The optimization objective function is:

[0111]

[0112] where λ = 0.01 is the sparsity regularization coefficient. The IPOPT solver is used for nonlinear programming, and the maximum number of iterations is set to 1000 times, with the convergence accuracy Δf ≤ 10 -8 . The case element vector x * optimally generated is input into the adversarial behavior simulation unit.

[0113] Furthermore, the simulation unit trains the suspect's anti-investigation strategy network based on the Proximal Policy Optimization (PPO) algorithm, and its reward function is defined as:

[0114] r(s,a) = w1·C 隐藏 -w2·R 暴露 +w3·S 合理

[0115] Variables and calculation relationships:

[0116] Evidence hiding degree, N 发现 Number of evidences discovered by the trainee, N 总 Total number of evidences in the case;

[0117] Exposure risk, calculated as the overlapping ratio of the trajectory and the monitoring area;

[0118] S 合理 = cos(θ): Rationality of behavior, where θ is the cosine similarity between the current behavior pattern and historical cases; w1 = 0.7, w2 = 0.3, w3 = 0.1 are fixed weight coefficients.

[0119] Specifically, the policy network architecture:

[0120] Input layer: Suspect status vector (including position coordinates, timestamp, executed actions);

[0121] Hidden layer: 3-layer fully connected network (256 - 128 - 64 nodes), with the activation function being Swish;

[0122] Output layer: The action probability distribution π(a|s) is generated by Softmax, and the action space includes 10 types of anti-investigation behaviors such as "transfer physical evidence", "forge traces", and "change route".

[0123] During training, the KL divergence is used to constrain the amplitude of policy update, with the threshold set to δ = 0.01 and the learning rate η = 3×10 -4 , and the batch size is set to 128. The trained policy network generates a dynamic case script (including evidence chain adjustment and suspect behavior sequence), and outputs to the multi-modal virtual-real fusion training module at most.

[0124] Furthermore, the cross-modal alignment engine part includes a text encoding component, an image encoding component, and a semantic alignment component. The text encoding component adopts the BERT-base model (12 layers, hidden layer dimension 768, number of attention heads 12) fine-tuned based on public security interrogation texts, with the input being text clues (interrogation transcripts, communication records) in the virtual case, the maximum sequence length being 512, and outputs a text feature vector The image encoding component adopts the ResNet-50 architecture (input resolution 224×224, number of channels 3), pre-trained on the dataset of involved images (5000 labeled images, 10 types of physical evidence), and outputs an image feature vector The semantic alignment component maps the image features to the text feature space through a learnable projection matrix W∈R768×2048W∈R768×2048, and the alignment loss function is defined as:

[0125]

[0126] where

[0127] λ = 0.8 is the balance coefficient,

[0128] p text = Softmax(φ text (s i )) / τ), p image = Softmax(Wφ image (v i )) / τ), and the temperature coefficient τ = 0.07.

[0129] During training, the Adam optimizer is used (learning rate 1×10-4, batch size 32), and the training termination condition is that the change rate of the loss function is <1% for 5 consecutive epochs. The aligned multi-modal data is output to the dynamic variable injector part through the PCIe4.0 interface.

[0130] Preferably, the dynamic variable injector part includes an environment modeling component and a diffusion model component. The environment modeling component uses a bidirectional long short-term memory network (BiLSTM), and the dynamic variable injector executes in the diffusion model:

[0131] v t = LSTM(d1,…,d T ), d t = [temperature t , humidity t , light intensity t

[0132] The input is a 24-hour sequence of environmental parameters (temperature ∈ [-20, 50] °C, humidity ∈ [0, 100]%, light intensity ∈ [0, 10 ^ 5] lux), normalized to the [-1, 1] interval. The BiLSTM hidden layer dimension is 128, the time step T = 24, and the environmental state vector is output The diffusion model component uses an improved DDPM (Denoising Diffusion Probabilistic Models), and the noise scheduling rule is:

[0133]

[0134] Furthermore, the noise injection process superimposes environmental variables:

[0135]

[0136] The noise prediction network ∈ θ uses a U-Net structure (basic number of channels 64, number of attention layers 3, residual connection), and the training loss is:​

[0137]

[0138] The optimizer is AdamW (learning rate 5×10-5, weight decay 0.01), the training batch size is 64, and the gradient clipping threshold is 1.0. The data after injecting the environment variables is transmitted to the interference clue generator part through the NVLink high-speed channel.

[0139] Preferably, the interference clue generator part includes a game strategy component and a generative adversarial network component. The game strategy component defines a two-player non-zero-sum game. The student strategy space A1 contains 5 types of detection actions, and the system strategy space A2 contains 3 types of interference patterns. The mixed Nash equilibrium is solved by the policy gradient method, and finally the mixing ratio is fixed at η = 0.7·x real +0.3·x fake . The generative adversarial network component adopts the WassersteinGAN architecture. The generator G inputs noise The network structure is a fully connected layer (256-128-64, LeakyReLU activation, slope 0.2), and outputs forged clues (d is consistent with the dimension of the real clue). The discriminator D is a 4-layer convolutional network (convolution kernel 3×3, stride 2, number of channels 64-128-256-512, spectral normalization constraint), and the loss function is:

[0140]

[0141] where γ = 10, The optimizer is Adam (β1 = 0.5, β2 = 0.9, learning rate 1×10 -4 ), and it is trained until the relative standard deviation of the discriminator loss < 5%. After the mixed data passes the quality verification (PSNR ≥ 30dB, SSIM ≥ 0.85), it is output to the mixed reality terminal through the DP1.4 interface

[0142] Preferably, the causal reinforcement learning evaluation module includes a temporal causal graph construction part, a causal contribution quantification part, and a meta-learning dynamic adaptation part. The temporal causal graph construction part receives the student operation sequence data output by the multi-modal virtual-real fusion training module. Exemplarily, the data includes action types ("retrieve surveillance", "fingerprint comparison"), timestamps (millisecond precision), and operation object identifiers. The temporal dependence features are extracted through a temporal convolutional network (TCN). Preferably, the TCN contains 4 residual blocks, each block consists of a dilated convolutional layer (dilation factor d = 1, 2, 4, 8), a weight normalization layer, and a ReLU activation layer. The convolutional kernel size is fixed at 3, and the output is a causal adjacency matrix M ∈ {0,1} N×N, where N is the number of operation nodes. Exemplarily, when the length of the operation sequence exceeds 100, it is truncated to N = 100 using a sliding window.

[0143] Exemplarily, the causal contribution quantification unit calculates the causal effect of each operation on the case-solving result based on the causal adjacency matrix. Further, counterfactual intervention operations do(a i = 1) and do(a i = 0) are defined, and counterfactual samples are generated through a generative adversarial network (GAN). The generator G is preferably a 3-layer fully connected network (input noise dimension 128, hidden layer 256 - 128, output dimension the same as the operation features, activation function is LeakyReLU, slope 0.2), and the discriminator D is a 2-layer convolutional network (convolution kernel 5×5, number of channels 32 - 64, spectral normalization constraint). The formula for calculating the causal contribution is:

[0144]

[0145] where K = 1000 is the number of Monte Carlo sampling times, and y ∈ [0, 1] represents the case-solving success rate.

[0146] Further, the counterfactual sample generation loss function is defined as:

[0147]

[0148] where λ = 0.1, the MMD kernel function is a Gaussian kernel (bandwidth σ = 1.0), and the input noise The training batch size is 64, the optimizer is Adam (β1 = 0.5, β2 = 0.9), until the discriminator loss volatility < 5%.

[0149] Further, the meta-learning dynamic adaptation unit adjusts the case generation parameters based on the causal contribution evaluation results. Preferably, the model-agnostic meta-learning (MAML) algorithm is adopted, and the meta-learning loss function is defined:

[0150]

[0151] where the inner loop learning rate α = 0.01, the outer loop learning rate β = 0.001, and the task distribution covers the evidence chain complexity C ∈ {1, 2, 3} (exemplarily, C = 3 corresponds to difficult cases with the number of evidence chain nodes ≥ 10). The number of parameter update iterations is 100, the optimizer is Adam (β1 = 0.9, β2 = 0.999), and the gradient clipping threshold is 1.0. The adjusted parameters are fed back to the non-linear constraint optimization unit of the case dynamic generation module through an AES-256 encrypted channel

[0152] The fully homomorphic encryption security module includes a key generation unit, a ciphertext operation unit, and a noise management unit. The key generation unit generates a key pair based on the Ring Learning with Errors (RLWE) problem. Exemplarily, a polynomial ring is defined where the modulus q = 2 32 - 1, and the polynomial degree n = 4096. The private key s is preferably sampled from a ternary distribution {-1, 0, 1} n , and the public key is constructed as p = [a, b = a·s + 2e], where noise (discrete Gaussian distribution, standard deviation σ = 3.2). Further, the noise sampling is implemented by the Knuth - Yao algorithm to ensure resistance to timing attacks, and the sampling time consumption ≤0.1 μs / coefficient.

[0153] Supports homomorphic operations on tensor data, adopts the CKKS encryption scheme, the polynomial ring dimension N = 8192, the modulus q = 2 32 - 1, and the specific operation rules are:

[0154]

[0155] where:

[0156] CPD represents the Canonical Polyadic Decomposition of the tensor, and the decomposition rank R = 32;

[0157] λ r is the decomposition coefficient, which is calculated by an iterative algorithm in the encrypted domain;

[0158] The operation error ≤1×10 -5 , meeting the accuracy requirements for investigating teaching data.

[0159] The ciphertext operation unit receives the plaintext data m ∈ {0, 1}, and generates the ciphertext c = (c0, c1) through the encryption algorithm. The encryption formula is:

[0160] c0 = a·r + 2e1 + mc1 = b·r + 2e2

[0161] where r ~ {-1, 0, 1} n , noise The ciphertext addition and multiplication operations are respectively defined as:

[0162]

[0163] Preferably, after the multiplication operation, the noise growth is reduced through the key exchange technology. The exchanged key k sw =(a sw , b sw = a sw ·s + 2esw ) are pre-generated and stored in the HSM, where After the exchange, the ciphertext noise level is reduced from O(η 2 ) to O(η), where η is the initial noise level.

[0164] The noise management unit monitors the ciphertext noise level in real time and defines a noise estimation function:

[0165] η(c) = ||c1 - s·c0|| ∞

[0166] When η(c) ≥ B (threshold B = 2 20 ), a re-encryption operation is triggered. Further, the re-encryption adopts a modulus switching technique:

[0167]

[0168] where the new modulus q ′ = 2 22 , and the noise reduction factor is 2 10 .

[0169] The key generation unit is connected to the encryption engine of the ciphertext operation unit through a Hardware Security Module (HSM, compliant with FIPS140-2 Level 3), and the public key is stored in the NVM of the encryption engine (the number of erase-write cycles ≥ 10^6 times, and the data retention period ≥ 10 years);

[0170] The ciphertext operation unit is connected to the FPGA computing unit of the noise management unit through a PCIe4.0×16 bus (bandwidth 64GB / s, latency ≤ 1μs) (model Xilinx Alveo U280, configured with 4 computing engines for parallel processing);

[0171] The output end of the noise management unit is fed back to the key generation unit through a TLS1.3 encryption channel (the key exchange algorithm is ECDH-SECP384R1, and the encryption suite is AES-256-GCM), and the key update period is 24 hours.

[0172] The public security service expansion interface module includes a protocol adaptation unit, a data conversion engine unit, a security authentication unit, and a traffic scheduling unit, and is dedicated to the public security AI-driven data generation-based investigation teaching and training system. The protocol adaptation unit includes a multi-protocol parsing component and a dynamic interface mapping component. Exemplarily, the protocol parsing component supports GB / T28181, Onvif, and an analog investigation instruction set (including 12 types of operations such as on-site investigation and evidence chain reasoning), and the interface mapping component converts heterogeneous data into a unified training format through a dynamic field mapping table (field matching rate ≥ 95%). The protocol parsing component preferably adopts a finite state machine model and defines state transition rules:

[0173] S t+1 = f(St , B t , C)

[0174] where S t is the current state (the number of states N = 24 covering the entire investigation process), B t is the input data stream, is the context vector (including case type, trainee level). The state transition function f is stored in the FPGA lookup table (LUT depth 8192), and the parsing speed ≥ 12 Gbps.

[0175] The data conversion engine department receives the output data of the protocol adaptation department and performs structured conversion and data enhancement. Define the data generation model:

[0176]

[0177] where the generator G is preferably a 3-layer fully connected network (input noise dimension 128, hidden layer 256 - 128, output dimension 256), used to generate simulated case data (suspect trajectory, physical evidence distribution). The discriminator D is a 2-layer CNN (convolution kernel 5×5, channels 32 - 64), and the loss function is:

[0178] L adv = E[logD(Y real )] + E[log(1 - D(G(z)))]

[0179] The generated data is compressed by Zigzag coding (compression rate ≥ 65%) and then input into the investigation teaching AI model.

[0180] The security authentication department includes a dynamic token generation component and a teaching permission control component. The dynamic token algorithm is:

[0181] T k = HMAC - SHA3 - 512(K master , Timestamp||SessionID)

[0182] where the master key K master is stored in the HSM (compliant with the GM / T 0054 - 2018 standard), and SessionID is a 32-byte globally unique identifier. The permission control is based on the RBAC - ABAC hybrid model, and the policy is defined as:

[0183]

[0184] where the weights of A1 (trainee role), A2 (training progress), and A3 (operation compliance) are w1 = 0.5, w2 = 0.3, w3 = 0.2, and the threshold 0.85 ensures that high-risk operations are only open to senior trainees.

[0185] The traffic scheduling department uses the adaptive weighted round-robin algorithm to allocate teaching requests, and the weight calculation is as follows:

[0186]

[0187] Where L j is the server node load (CPU utilization rate ≤ 80%), α = 1.5 is the attenuation factor, and T task is the current task complexity, (evidence chain length × 0.3 + time constraint × 0.7), and T avg is the average task time-consuming of the system.

[0188] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The data generation-based investigation teaching and training system driven by public security AI is characterized in that, It includes the following modules: A case dynamics generation module, which is used to receive the original data of the public security case database, generate logically closed virtual cases through tensor decomposition, and transmit the generated case data to the multi-modal virtual-real fusion training module; A multi-modal virtual-real fusion training module, which is used to receive the virtual case data, realize the interactive training of multi-modal clues through mixed reality technology, dynamically inject interfering clues, and transmit the trainee operation data to the causal reinforcement learning evaluation module; A causal reinforcement learning evaluation module, which is used to receive the trainee operation data, construct a temporal causal graph and quantify the causal contribution degree, and generate an evaluation report to feedback to the case dynamics generation module to adjust the case generation parameters; A fully homomorphic encryption security module, which is used to encrypt and decrypt the transmission data between the case dynamics generation module, the multi-modal virtual-real fusion training module, and the causal reinforcement learning evaluation module in real time; A public security business extension interface module, which receives real-time case feature data from the public security actual combat platform and imports it into the case dynamics generation module, and at the same time exports the trained case plan to the police command system.

2. The public security AI-driven data generation-based investigation teaching and training system according to claim 1, characterized in that, The case dynamics generation module includes: A case element preprocessing unit, which cleans the original data of the public security case database and extracts suspects, physical evidence, time, space entities and their relationships; Case element tensor encoding unit, encoding the preprocessed entity relationship into a fourth-order tensor where I, J, K, and L represent the suspect attribute, physical evidence type, time slice, and spatial grid dimension respectively, and perform Tucker decomposition to extract potential association patterns; A non-linear constraint optimization unit, which constructs and solves an investigation logic constraint equation set based on the decomposition result to generate a virtual case with a closed evidence chain; An adversarial behavior simulation unit, which simulates the anti-investigation behavior of suspects through reinforcement learning to generate a dynamically evolving case script.

3. The public security AI-driven data generation-based investigation teaching and training system according to claim 1, characterized in that, The multi-modal virtual-real fusion training module includes: A cross-modal alignment engine, which performs semantic alignment on text clues and image clues to generate consistent multi-modal data; A dynamic variable injector, which superimposes spatio-temporal environment variables during the generation process of the diffusion model; An interfering clue generator, which mixes real clues and interfering clues according to a preset ratio based on game theory strategies; the data processed by the cross-modal alignment engine is input into the dynamic variable injector, and the data after injecting the environment variables is adjusted by the interfering clue generator and output to the training scenario.

4. The public security AI-driven data generation-based investigation teaching and training system according to claim 1, characterized in that, The causal reinforcement learning evaluation module includes: A temporal causal graph construction unit, which extracts the causal dependence relationship of the trainee operation sequence through a temporal convolutional network; A causal contribution degree quantification unit, which calculates the causal effect of the operation on the case-solving result; A meta-learning dynamic adaptation unit, which dynamically adjusts the case generation difficulty based on the evaluation result.

5. The data generation-based investigation teaching and training system driven by public security AI according to claim 2, characterized in that The case element tensor encoding unit performs Tucker decomposition to satisfy; Where: is the core tensor, representing the potential cross-dimensional correlations; Factor matrices for suspect attributes and physical evidence types, respectively; Factor matrices for time and space respectively.

6. The public security AI-driven data generation-based investigation teaching and training system according to claim 2, characterized in that, The equation set solved by the non-linear constraint optimization unit is: Where: x i ∈ [0, 1]: Activation probability of the i-th element in the case element vector; α i ∈ [0, 1]: Weight coefficient of evidence, calculated from the occurrence frequency of similar physical evidence in historical cases; β = 0.8: The evidence chain closure threshold, indicating that at least 80% of the key evidence associations need to be satisfied; Spatial-temporal correlation matrix, element a ij Indicates the spatial accessibility between positions i and j; A reference vector, representing a standard spatio-temporal distribution pattern; ∈ = 0.1: The tolerance parameter, allowing the spatio-temporal distribution deviation not to exceed 10%.

7. The data generation-based investigation teaching and training system for public security driven by AI according to claim 3, characterized in that, The loss function of the cross-modal alignment engine is: Where: λ = 0.8 is the balance coefficient, p text = Softmax(φ text (s i ) / τ), p image = Softmax(Wφ image (v i ) / τ), temperature coefficient τ = 0.

07.

8. The AI-driven data generation-based investigation teaching and training system for public security according to claim 1, characterized in that, The dynamic variable injector executes in the diffusion model: v t = LSTM(d1,…,d T ), d t = [Temperature t , Humidity t , Illumination t ​ Where: The dimension of the hidden layer of the LSTM network is 128, and the time step T = 24; The noise scheduling parameter α of the diffusion model t Decays according to the cosine rule.

9. The public security AI-driven data generation-based investigation teaching and training system according to claim 4, wherein The causal contribution degree quantification unit satisfies: Where: Is an indicator function that outputs 1 when the operation exists and 0 otherwise; Counterfactual samples Generated by GAN, and the generator is a 3-layer MLP.

10. The public security AI-driven data generation-based investigation teaching and training system according to claim 1, wherein The fully homomorphic encryption security module performs the following operations: Encrypted Tensor Operations: Homomorphic encryption is performed on tensor data in case generation and evaluation, satisfying: Among them, the CPD decomposition rank R = 32, the encryption scheme is CKKS, and the polynomial modulus degree N = 8192; Edge-Cloud Collaboration: Edge nodes process real-time interactive data, and the cloud executes offline model training; Dynamic Key Management: Rotate the encryption key every 24 hours and protect historical data through a forward security protocol.

Images