Event anti-cheating and precise timing method and system based on multi-modal data fusion

Abstract

The application relates to the technical field of event anti-cheating, and particularly discloses an event anti-cheating and precise timing method and system based on multi-modal data fusion, which comprises the following steps: a federal computing network with a global time reference server as the core is constructed, and independent federal node identities and unique blockchain identity identifiers are allocated to each punch-in point edge computing unit and a player wearable device; and multi-modal event data, including radio frequency punch-in data, is collected within a preset time slice. In actual event application, the technical scheme provided by the application realizes irreversible desensitization and splicing of multi-modal event data into a feature tensor at the edge of a punch-in point, preliminary cheating determination is completed by relying on a local individualized model, and then only encrypted parameter increments and suspicious feature vectors are uploaded, so that the full amount of original multi-modal data is effectively returned, and the transmission delay problem under limited bandwidth outdoors is solved.

Description

Technical Field

[0001] This invention relates to the field of anti-cheating technology for sports events, and in particular to a method and system for anti-cheating and accurate timing of sports events based on multimodal data fusion. Background Technology

[0002] Existing sports event timing and anti-cheating systems mainly rely on radio frequency identification tags or single positioning devices. When athletes pass through checkpoints, the system collects single-modal check-in data and transmits it directly back to the central server. The central server then uses preset fixed rules to determine whether there are any violations such as running on behalf of others or deviating from the course.

[0003] However, single data points are extremely easy to forge or transfer. Introducing multimodal data, such as physiological and visual data, to improve the accuracy of anti-cheating detection is an inevitable trend, but this raises the following problems:

[0004] Under limited outdoor communication bandwidth, synchronously transmitting massive amounts of multimodal raw data back to the central server can lead to severe network congestion and decision-making delays; while performing isolated data processing only at the checkpoint will result in an extremely high frequency of misjudgments due to the lack of global data comparison and spatiotemporal correlation analysis.

[0005] Existing technologies cannot simultaneously guarantee high-precision cheating detection of multimodal data while resolving the contradiction between high-concurrency data transmission latency and global overall judgment. Summary of the Invention

[0006] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the purpose of this invention is to propose a method and system for preventing cheating and ensuring accurate timing in sports events based on multimodal data fusion, effectively eliminating the risk of cheating and misjudgment in complex sports environments.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for preventing cheating and ensuring accurate timing in sports events based on multimodal data fusion, comprising:

[0008] Construct a federated computing network with a global time base server at its core, and assign independent federated node identities and unique blockchain identity identifiers to edge computing units at each checkpoint and wearable devices of athletes;

[0009] Multimodal event data is collected within a preset time slice, including radio frequency check-in data, athlete physiological data, and visual monitoring data.

[0010] The multimodal competition data is irreversibly desensitized at the data acquisition end, and then spliced ​​together after spatiotemporal alignment to generate a unified multidimensional spatiotemporal feature tensor.

[0011] Based on the unified multidimensional spatiotemporal feature tensor, local personalized model incremental training is performed in the edge computing unit of the check-in point, and the Mahalanobis distance between the current feature vector and the normal feature distribution is calculated to generate a preliminary cheating judgment result on the edge side; when the preliminary cheating judgment result on the edge side is abnormal, the corresponding feature vector is extracted as a suspicious feature vector.

[0012] The parameter increments of the local personalized model are encrypted and differentially privacy processed before being uploaded. The updated global benchmark model is generated and distributed through the encrypted domain federation aggregation of the central server.

[0013] The suspicious feature vectors are input into the global benchmark model, and the final judgment is generated by combining the cross-node spatiotemporal cross-validation results, and the athlete's global score is updated simultaneously.

[0014] To achieve the above objectives, a second aspect of the present invention proposes a competition anti-cheating and accurate timing system based on multimodal data fusion, comprising:

[0015] The network construction and identity allocation module is used to build a federated computing network with a global time base server as the core, and to assign independent federated node identities and unique blockchain identity identifiers to the edge computing units at each checkpoint and the wearable devices of athletes.

[0016] The data acquisition and tensor generation module is used to acquire multimodal event data within a preset time slice. The multimodal event data includes radio frequency check-in data, athlete physiological data, and visual monitoring data. The module performs irreversible desensitization processing on the multimodal event data at the data acquisition end and splices it together after spatiotemporal alignment to generate a unified multidimensional spatiotemporal feature tensor.

[0017] The edge-side training and preliminary judgment module is used to perform local personalized model incremental training on the edge computing unit of the check-in point based on the unified multidimensional spatiotemporal feature tensor, and calculate the Mahalanobis distance between the current feature vector and the normal feature distribution to generate an edge-side preliminary cheating judgment result; when the edge-side preliminary cheating judgment result is abnormal, the corresponding feature vector is extracted as a suspicious feature vector.

[0018] The encrypted domain federated aggregation and update module is used to encrypt and perform differential privacy processing on the parameter increments of the local personalized model before uploading it, and generate an updated global benchmark model through encrypted domain federated aggregation of the central server and distribute it.

[0019] The global collaborative verification and score generation module is used to input the suspicious feature vector into the global benchmark model, combine the cross-node spatiotemporal cross-verification results to generate the final judgment conclusion, and simultaneously update the athlete's global score.

[0020] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for preventing cheating and accurate timing in competitions based on multimodal data fusion.

[0021] In practical competition applications, the technical solution provided by this invention irreversibly desensitizes multimodal competition data at the edge of checkpoints and concatenates it into feature tensors. Preliminary cheating detection is then performed using a local personalized model. Subsequently, only encrypted parameter increments and suspicious feature vectors are uploaded, effectively avoiding the full back transmission of the original multimodal data and resolving the transmission latency problem under limited outdoor bandwidth. Furthermore, the central server uses a global benchmark model generated by encrypted domain federated aggregation to perform secondary analysis on suspicious feature vectors and introduces a cross-node spatiotemporal cross-verification mechanism to compensate for the blind spots in local edge-end judgments. While ensuring that the privacy data of the participants does not leave the edge, it successfully balances low-latency edge processing with global overall verification, effectively eliminating the risk of cheating and misjudgment in complex competition environments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the anti-cheating and accurate timing method for competitions based on multimodal data fusion provided by the present invention.

[0023] Figure 2 This is the error convergence curve of the two-way time synchronization protocol compensation algorithm in the anti-cheating and accurate timing method for competitions based on multimodal data fusion provided by this invention;

[0024] Figure 3 This is a comparison diagram of the athlete's original PPG waveform and FFT frequency domain feature extraction in the anti-cheating and accurate timing method for competitions based on multimodal data fusion provided by this invention;

[0025] Figure 4 This invention provides a method for preventing cheating and ensuring accurate timing in competitions based on multimodal data fusion, which includes the distribution and confidence elliptic plot of multidimensional feature anomaly detection on the edge side based on Mahalanobis distance.

[0026] Figure 5 This invention provides a method for preventing cheating and ensuring accurate timing in competitions based on multimodal data fusion, which includes an adaptive calibration surface plot of node aggregation weights based on the tanh function under local environmental abrupt changes.

[0027] Figure 6 This invention provides a heatmap of the spatial mapping between the track micro-digital elevation model (DEM) and terrain resistance characteristic values ​​in a race anti-cheating and accurate timing method based on multimodal data fusion.

[0028] Figure 7This is a comparison curve of the cold start loss function of the node-level initial model based on knowledge distillation in the anti-cheating and accurate timing method for competitions based on multimodal data fusion provided by this invention.

[0029] Figure 8 This is a schematic diagram illustrating the implementation of the anti-cheating and accurate timing system for competitions based on multimodal data fusion provided by the present invention.

[0030] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following description, with reference to the accompanying drawings, describes an embodiment of the present invention of a competition anti-cheating and accurate timing system, method, and electronic device based on multimodal data fusion.

[0033] Example 1:

[0034] like Figure 1 As shown, this embodiment of the invention provides a method for preventing cheating and ensuring accurate timing in sports events based on multimodal data fusion.

[0035] In existing sports events, especially long-distance outdoor endurance races such as 100km trail running and mountain marathons, the actual physical environment is usually extremely complex, accompanied by objective limitations such as unstable communication networks, drastic terrain undulations, and variable microclimates.

[0036] This embodiment will provide a detailed description of the technical solution in conjunction with the above-mentioned practical application scenarios. The entire technical solution can be divided into the following steps:

[0037] Step 1: Construct a federated computing network and bind it to the physical layer isolation.

[0038] Specifically, before the competition begins, the system first needs to build the underlying basic network architecture. In this embodiment, a federated computing network with a global time base server at its core is constructed, and independent federated node identities and unique blockchain identity identifiers are assigned to the edge computing units at each checkpoint and the wearable devices of athletes.

[0039] For example, the edge computing units at checkpoints can be deployed at various timing checkpoints (CP points) along the track. Their hardware architecture can utilize industrial-grade edge computing gateways equipped with neural processing units (NPUs). Athletes' wearable devices are smart terminals worn on their wrists or chests. To achieve unified scheduling and timing synchronization of these discrete nodes, the system generates a unique global time base server within the cluster through a distributed lock election mechanism. Considering the potential for network outages at outdoor base stations, the aforementioned distributed lock election mechanism employs the Raft consensus algorithm, which features high availability. It is also important to note that the lock lifecycle of this global time base server is set to a first preset duration and periodically re-elects. This first preset duration can be set, for example, to 30 seconds. If the current master node fails, other candidate nodes within the cluster will initiate a new round of voting competition after the first preset duration expires, ensuring the continuity of the time base service.

[0040] To ensure high accuracy in event timing, the system uses a two-way time synchronization protocol compensation algorithm to correct network latency, constraining the time synchronization error accuracy of all independent federated nodes to within a second preset duration. This second preset duration is strictly set to 1 millisecond. The aforementioned two-way time synchronization protocol compensation algorithm calibrates the local clock by calculating the time offset, and the formula for calculating the time offset is as follows:

[0041] ;

[0042] in, This represents the time offset of the edge computing unit clock relative to the global time base server clock; This represents the local timestamp of the synchronization request message sent by the edge computing unit to the global time base server. This indicates the base timestamp at which the global time base server received the synchronization request message; The base timestamp that indicates the synchronization response message sent by the global time base server; This indicates the local timestamp at which the edge computing unit received the synchronization response message.

[0043] This algorithm eliminates errors caused by asymmetric delays in uplink and downlink network transmission, ensuring the rigor of all attendance records in the time dimension.

[0044] like Figure 2 As shown in the figure, the horizontal axis represents the number of network interaction rounds, in seconds, and the vertical axis represents the time synchronization error, in milliseconds.

[0045] The blue node synchronization error curve in the figure illustrates that when an edge computing unit initially connects to the federated computing network, its initial synchronization error reaches 35.2 milliseconds due to the asymmetric delay in transmission from the outdoor base station network. As the bidirectional time synchronization protocol compensation algorithm is executed periodically, the blue curve representing the error exhibits a rapid, exponentially decaying waveform transformation.

[0046] After four rounds of network interaction, the time synchronization error has dropped significantly to 2.4 milliseconds, and further to 0.85 milliseconds in the fifth round of network interaction, successfully entering the error constraint boundary of ±1 millisecond as indicated by the red dashed line in the figure.

[0047] During the subsequent normalized synchronization process up to the 15th network interaction, despite the outdoor network jitter, the blue node synchronization error curve always oscillated slightly around 0 milliseconds, and all data points were strictly limited within the red error constraint boundary, with a maximum reverse deviation of -0.42 milliseconds.

[0048] This data distribution and curve convergence trend prove that the system can eliminate the error caused by the asymmetric delay of uplink and downlink network transmission, and constrain the physical clock error within a preset range of 1 millisecond, thus laying a reliable time reference for the unified spatiotemporal alignment of multimodal data across the entire network.

[0049] Optionally, before collecting multimodal event data, a physical layer isolation binding operation is performed on the device. Specifically, this includes: physically binding the athlete's wearable device, which integrates a power-off self-destruct encryption unit, to the unique blockchain identity; and having the global scheduling center distribute the same-source key to the high-speed volatile static random access memory (SRAM) of the power-off self-destruct encryption unit. When an unauthorized read command is detected, the power supply to the memory is instantly cut off to destroy the same-source key. In practical applications, unauthorized read commands typically originate from cheaters attempting to extract the key from the device using hardware probes to forge attendance data. The high-speed volatile static random access memory (SRAM) allows the charge stored in the transistors to dissipate within microseconds after power failure, ensuring that the same-source key cannot be physically stolen, thus guaranteeing the unforgeability of the node's identity from the hardware level.

[0050] Step 2: Construction of the cold start mechanism for the federated model.

[0051] Because the physical environment of each event is different, directly using an uninitialized general model will lead to a high misjudgment rate in the early stages of the event. Therefore, in this embodiment, the federated computing network is also configured with a federated model cold start mechanism.

[0052] Specifically, the system acquires macro-environmental characteristic parameters of the current event, including altitude and ambient temperature. In practical event applications, altitude directly affects the oxygen content of the air, thus significantly altering the athlete's baseline heart rate and blood oxygen consumption rate; ambient temperature directly affects the athlete's heat dissipation efficiency and pacing strategy. Based on a longitudinal federated transfer learning algorithm, the system matches and extracts basic knowledge weights from a historical event federated model library according to the macro-environmental characteristic parameters, generates an initial global baseline model for the current event, and distributes it.

[0053] For example, the longitudinal federated transfer learning algorithm can pre-set the weights of the network neurons in the current initial global benchmark model by extracting public representation space mapping relationships under similar historical environments, such as past high-altitude and low-temperature events, while ensuring that historical event data does not leave the domain. This allows the model to have prior knowledge of the normal physiological fluctuation range under the current specific altitude and temperature environment even before the event generates a large amount of real-time data, effectively avoiding algorithmic decision oscillations caused by data scarcity during the cold start phase.

[0054] Step 3: Collection of multimodal event data and generation of spatiotemporal feature tensors.

[0055] During the event, the system collects multimodal event data within preset time slices. This multimodal event data includes radio frequency check-in data, athlete physiological data, and visual monitoring data. Irreversible anonymization processing is performed on the multimodal event data at the data acquisition end, and after spatiotemporal alignment, the data is spliced ​​to generate a unified multidimensional spatiotemporal feature tensor.

[0056] Specifically, the above process involves concurrent processing across multiple dimensions. First, for a first preset time slice, the radio frequency attendance data is converted into a first preset dimension radio frequency spatiotemporal feature tensor containing timestamps, locations, and kinematic information. This first preset time slice is, for example, set to 10 milliseconds, and the first preset dimension is, for example, set to 64 dimensions. These 64-dimensional features are filled with discrete motion physical quantities such as triaxial acceleration, triaxial angular velocity, and absolute latitude and longitude coordinate offsets, smoothed by Kalman filtering.

[0057] Secondly, a second preset-dimensional physiological feature tensor is locally extracted and generated on the athlete's wearable device, and a third preset-dimensional visual spatial feature tensor is extracted and generated on the edge computing unit at the check-in point. For example, the second preset dimension is set to 128 dimensions, and the physiological feature tensor is extracted from the raw waveform data collected by the photoplethysmography (PPG) sensor built into the athlete's wearable device via Fast Fourier Transform on the device's internal microcontroller, including frequency domain features such as heart rate variability and respiratory rate. The third preset dimension is set to 256 dimensions, and the visual spatial feature tensor is extracted from the raw video stream collected by the high-definition camera at the check-in point via convolutional layers of a lightweight residual neural network (ResNet) deployed on the edge computing unit.

[0058] like Figure 3 As shown in the attached figure, the waveform transformation process of the signal is intuitively presented through the comparison of the upper and lower regions. The upper part of the figure shows the time-domain waveform acquired by the sensor in real time, with the horizontal axis representing time in seconds and the vertical axis representing amplitude in millivolts.

[0059] The blue raw photoplethysmogram curve in the figure exhibits a continuous and complex fluctuation state, including low-frequency respiratory baseline drift and high-frequency environmental interference signals. This represents the raw sampled data, which contains extremely high personal physiological characteristic recognition before desensitization processing. The system performs a Fast Fourier Transform on this time-domain waveform through the device's internal microcontroller unit, converting it into the frequency domain distribution state shown in the lower half.

[0060] The lower half of the graph shows frequency on the horizontal axis (in Hertz) and frequency domain amplitude on the vertical axis (in millivolts). The red Fast Fourier Transform frequency domain characteristic curve, representing the conversion result, exhibits significant sharp peaks around 0.3 Hz and 2.5 Hz. These two peaks objectively reflect the athlete's true respiratory rate and heart rate distribution characteristics under high-intensity exercise load.

[0061] This waveform transformation operation, which converts complex and continuous time-domain fluctuation curves into discrete frequency-domain peaks, demonstrates that the system can effectively strip away and overwrite and destroy personal biometric privacy information in the original waveform, and accurately extract the core biomechanical frequency-domain data used to construct the 128-dimensional physiological feature tensor. This multimodal data desensitization processing result provides high-confidence data support for subsequent edge-side local personalized model training and preliminary cheating judgment based on Mahalanobis distance.

[0062] It is also important to note that irreversible desensitization processing means that, regardless of whether it is the original pulse waveform or the original facial image, after extracting it into an abstract high-dimensional floating-point feature tensor, the original sampling file is immediately overwritten and destroyed in memory, effectively preventing the risk of facial recognition or medical privacy leakage.

[0063] Subsequently, the system performs forced spatiotemporal alignment and concatenates the physiological feature tensor, the visual spatial feature tensor, and the radio frequency spatiotemporal feature tensor according to their corresponding timestamps, generating a unified multidimensional spatiotemporal feature tensor with a fourth preset dimension. The system then calculates the secure hash value of this tensor and stores it on the blockchain. The fourth preset dimension is the sum of the dimensions of the three tensors mentioned above, i.e., 448 dimensions. Forced spatiotemporal alignment employs linear interpolation and forward filling algorithms for time windows to address data misalignment caused by inconsistent sampling frequencies of different sensors. The generated 448-dimensional abstract vector represents the athlete's holographic state within a specific 10-millisecond slice. By calculating a secure hash value, such as using the SHA-256 algorithm, and storing the data on the blockchain, the tamper-proof evidence storage of this multimodal data is ensured during subsequent arbitration.

[0064] Step 4: Local personalized model training and preliminary cheating detection on the edge side.

[0065] After acquiring the structured tensor data, the system performs local personalized model incremental training on the edge computing unit at the check-in point based on the unified multidimensional spatiotemporal feature tensor, and calculates the Mahalanobis distance between the current feature vector and the normal feature distribution to generate a preliminary cheating judgment result at the edge. When the preliminary cheating judgment result at the edge is abnormal, the corresponding feature vector is extracted as a suspicious feature vector.

[0066] Specifically, the local personalized model adopts an autoencoder network architecture. As the competition progresses, the edge computing unit uses the continuously flowing in of athletes' real-time 448-dimensional tensors to incrementally update the parameters of the autoencoder through backpropagation, enabling it to dynamically fit the unique feature distribution pattern of each athlete in the current stage of the competition.

[0067] During anomaly detection, the system uses the Mahalanobis distance algorithm for initial judgment. The Mahalanobis distance between the current feature vector and the normal feature distribution is calculated using the following formula:

[0068] ;

[0069] in, Indicates the current time Local Mahalanobis distance of the feature vectors; This represents the feature vector corresponding to the unified multidimensional spatiotemporal feature tensor of the current input; This represents the normal feature mean vector output by the local personalized model; The covariance matrix representing local normal features; Represents the transpose of the characteristic deviation vector; It represents the inverse of the covariance matrix of the local normal features.

[0070] In practical applications, compared with the traditional Euclidean distance, Mahalanobis distance, by introducing the inverse matrix of the covariance matrix, effectively eliminates the interference caused by the inconsistency of dimensions between multimodal features and the high correlation of features (such as the natural positive correlation between an athlete's running cadence and heart rate), greatly improving the scientific nature of the definition of abnormal boundaries.

[0071] The specific rules for generating preliminary cheating judgment results on the edge side are as follows:

[0072] like Then it is judged as normal, among which Local standard deviation, The first preset multiple threshold;

[0073] like Then it is determined to be abnormal and marked as the suspicious feature vector, where The second preset multiple threshold;

[0074] like It is then determined to be high-risk and a temporary interception command is generated, and the first preset multiple threshold is set. Strictly less than the second preset multiple threshold For example, when a player's feature deviation is within a slightly abnormal range, such as a trajectory deviation caused by taking a shortcut, but the physiological heart rate has not yet dropped significantly, the system will identify the player's current feature vector as a suspicious feature vector and block it for further verification.

[0075] If the distance is extremely large, such as when the indicator shows that the athlete's speed reaches 60 kilometers per hour and their heart rate is in an extremely flat state, it is obviously cheating by vehicle, and an alarm will be triggered directly at the current checkpoint to intercept the cheating.

[0076] like Figure 4 As shown in the figure, the horizontal axis represents step frequency in steps per minute, and the vertical axis represents heart rate in beats per minute.

[0077] The dense blue dots in the figure represent the normal feature distribution of the local personalized model output. These data objectively show the natural positive correlation between the athlete's running cadence and heart rate, with the data center roughly concentrated at the cadence of 160 and heart rate of 150.

[0078] Around the center of this normal data distribution, the graph expands outwards into two layers of sloping confidence ellipses. The inner yellow dashed line represents the first multiple threshold boundary, and the outer red dashed line represents the second multiple threshold boundary. Because the Mahalanobis distance calculation mechanism introduces the inverse of the covariance matrix, it effectively eliminates the interference of positive correlations between features. This allows the two threshold boundaries to accurately match the trend of the true features, thus forming sloping ellipses rather than regular circles.

[0079] When the input features, as shown by the yellow suspicious feature vector scatter points, fall between the first multiple threshold boundary and the second multiple threshold boundary, for example, when the cadence reaches 185 but the heart rate is only 130, the system determines that it deviates from the normal physiological synchronization features, classifies it into the slightly abnormal interval, and extracts and reports it.

[0080] When the input features, as shown by the red high-risk feature vector scatter points, significantly exceed the second multiple threshold boundary on the outer side, such as a step frequency as high as 190 and a heart rate abnormally flat to 100, this data performance that seriously violates the normal biomechanical laws will be identified by the system as high-risk cheating behaviors such as taking a car to run instead of running, and a temporary interception command will be generated immediately.

[0081] The attached figure visually demonstrates that the system can accurately define abnormal boundaries and achieve high-precision cheating interception by relying solely on multidimensional correlation analysis at the edge.

[0082] Step 5: Encryption of model parameters, differential privacy processing, and federated aggregation of ciphertext domains.

[0083] To achieve global knowledge sharing and prevent privacy through reverse inference, the system encrypts and performs differential privacy processing on the parameter increments of the local personalized model before uploading it. The updated global benchmark model is then generated and distributed through the encrypted domain federation aggregation of the central server.

[0084] Specifically, the parameter increments of the local personalized model are extracted at preset time intervals (e.g., every 5 minutes), and the parameter increments are ciphertext-transformed using a homomorphic encryption algorithm (such as the CKKS fully homomorphic encryption scheme). Homomorphic encryption allows specific algebraic operations while maintaining data encryption, placing the central server in a state where it can only perform collaborative computations in the ciphertext domain and cannot obtain the plaintext features of the data.

[0085] Simultaneously, the standard deviation of Gaussian noise is dynamically matched based on the sensitivity of the parameter increment, and the Gaussian noise is injected into the encrypted parameter increment based on an adaptive differential privacy mechanism, while setting a differential privacy budget. ,in This is a preset privacy and security threshold.

[0086] In theoretical applications, the privacy budget determines the intensity of the noise, and the set threshold strictly limits the methods used by cheating organizations. Even if a cheating organization intercepts the model parameter packet, it cannot reverse engineer the player's single-point spatiotemporal trajectory or pathological-level physiological characteristics through member inference attacks. Subsequently, the incremented ciphertext parameters with added noise are uploaded to the central server via an isolated encrypted channel.

[0087] On the central server, the model quality score for each node is calculated based on local verification metrics, such as the cross-entropy loss value on the local test set, which are synchronously uploaded by each independent federated node with the incremental ciphertext parameters. While the central server remains undecrypted, an incremental federated averaging algorithm is used to weighted aggregate the incremental ciphertext parameters uploaded by all independent federated nodes.

[0088] In the weighted aggregation process, a single-node weight factor is defined. The The data volume updated at the corresponding node and the model quality score are positively correlated. The central server performs a weighted aggregation calculation on the encrypted parameters, and the aggregation formula is as follows:

[0089] ;

[0090] in, Indicates the first The global baseline model parameters after each round of updates; Indicates the first The current global baseline model parameters; This represents the total number of independent federation nodes participating in this round of aggregation; Indicates the first The single-node weighting factor of each independent federation node; Indicates the first The number of independent federation nodes in the 1st The increment of ciphertext parameters uploaded in each round, after homomorphic encryption and noise addition. This formula ensures that high-quality, high-throughput check-in points contribute more to the global model, accelerating model convergence.

[0091] Optionally, the system further filters out abnormal parameter packets whose model quality scores are lower than a preset defense threshold, generates updated global baseline model parameters, and encrypts and distributes them to each independent federated node for parameter fusion with the local personalized model. This filtering mechanism aims to defend against malicious nodes that may inject abnormal parameters during the competition, such as cheaters hijacking a checkpoint and deliberately uploading polluting parameters to paralyze the anti-cheating system.

[0092] Step Six: Global Collaborative Verification and Generation of Final Judgment Conclusion.

[0093] Relying solely on single-point edge computing is insufficient to detect systemic cheating across multiple sites (such as proxy running). Therefore, the system inputs the suspected feature vectors into the global benchmark model, combines the cross-node spatiotemporal cross-validation results to generate a final judgment, and simultaneously updates the athlete's global performance.

[0094] Specifically, the global Mahalanobis distance between the suspected feature vector and the normal feature distribution of the global benchmark model is calculated. ,in This represents the suspicious feature vector. The calculation logic and formula for its global Mahalanobis distance are similar to the local judgment structure described above, the difference being that the evaluation benchmark here is a global distribution that integrates knowledge from all checkpoints.

[0095] To establish a robust chain of evidence, the system triggers a cross-node spatiotemporal cross-verification mechanism. This mechanism retrieves historical feature vectors from a predetermined number of adjacent checkpoints (e.g., three adjacent checkpoints) before and after the corresponding athlete, as well as parallel feature vectors from related athletes within a predetermined spatial radius (e.g., within a 100-meter radius), for spatiotemporal consistency comparison. In actual business logic, this spatiotemporal consistency comparison is used to verify whether the time taken for the athlete to reach the current checkpoint from the previous checkpoint violates human physiological limits, and whether the athlete's characteristics significantly deviate from the decay patterns of other real athletes in the same pace group.

[0096] The final decision logic is set as follows:

[0097] like Furthermore, the spatiotemporal consistency comparison passed, and it was determined to be normal. The first global distance threshold;

[0098] like If the spatiotemporal consistency comparison conditions are not fully met, a review instruction will be generated, in which... This is the second global distance threshold;

[0099] like Furthermore, the spatiotemporal consistency comparison condition was not met, thus it was determined to be cheating; and the first global distance threshold... Strictly less than the second global distance threshold .

[0100] Data that generates a pending review instruction will be pushed to the event organizing committee's large-screen terminal, where referees will conduct manual verification using video surveillance and other auxiliary means; if cheating is determined, the athlete's electronic timing score will be directly revoked.

[0101] Step 7: End-to-end blockchain audit mechanism.

[0102] To ensure the credibility of the judgment and meet the requirements of subsequent judicial audits, the method also includes a blockchain audit mechanism that runs through the entire process.

[0103] Specifically, the system packages the data hash values, model parameter incremental hash values, and final judgment conclusions within the time slice into a transaction set at preset block generation intervals. The transaction set is verified based on the Practical Byzantine Fault Tolerance (PBFT) consensus mechanism and written into the consortium blockchain's distributed ledger to generate an immutable traceability audit report. In actual network deployment, the consortium blockchain nodes are jointly maintained by the event organizer, an independent timing agency, and a third-party notary institution. The PBFT consensus mechanism requires that more than two-thirds of the ledger nodes confirm the legality of a transaction before a block can be written, ensuring that any unilateral backend modifications cannot take effect on the blockchain.

[0104] After the competition ends, the system de-identifies and extracts the features of the current global benchmark model, stores them in the federated model library, and automatically executes the destruction procedure for temporary data at edge nodes. This not only accumulates valuable knowledge assets for hosting similar competitions in the future (which can be reused through the federated model cold start mechanism), but also follows strict data lifecycle management standards, protecting the privacy rights of all participants.

[0105] Compared with the prior art, the embodiments of the present invention have the following technical effects: the existing single centralized anti-cheating system faces network congestion caused by the back transmission of raw high-dimensional data from the front end, as well as the technical conflict of serious lag and extremely high false judgment rate caused by simply relying on rule judgment.

[0106] This invention, through the introduction of edge-side desensitization fusion of multimodal spatiotemporal feature tensors, local dynamic anomaly screening based on Mahalanobis distance, and federated aggregation computation of ciphertext domains based on homomorphic encryption, constitutes a two-level collaborative network between the edge and the cloud. This solution not only effectively reduces reliance on weak outdoor communication bandwidth and avoids timing and statistics lag caused by network latency, but also significantly improves the accuracy of intercepting complex cheating methods, such as substitution, use of transportation, and shortcuts, through a cross-node spatiotemporal cross-verification mechanism, without exposing any athletes' original physiological image privacy data. This provides a highly reliable and credible engineering practice foundation for the fair operation and data security management of large-scale events.

[0107] Example 2:

[0108] In the technical solution provided in Embodiment 1 above, when the central server performs encrypted domain federated aggregation, it constructs an effective security defense mechanism for filtering abnormal parameter packets by comparing the model quality scores of each independent federated node with preset defense thresholds. The underlying logic of this mechanism is based on the assumption that the data presents independent and identical distributions, and it can reliably intercept data tampering or model parameter pollution attacks initiated by malicious cheaters.

[0109] However, in real-world applications such as 100km mountain trail races and extreme high-altitude endurance races, the course often traverses multiple microclimate zones and complex terrains. When a section of the race encounters sudden extreme weather events, such as torrential rain, hail, or a sharp drop in temperature, all participating athletes passing through the checkpoints in that area will inevitably experience highly consistent and drastic changes in their physiological load characteristics (such as abnormally high heart rate and decreased blood oxygen saturation) and kinematic characteristics (such as a sharp decrease in stride frequency and increased trajectory deviation).

[0110] This shift in group data distribution caused by abrupt changes in the objective physical environment can drive edge computing units at checkpoints in the region to generate model parameter increments that deviate significantly from the macro-global benchmark model during local personalized model training. Under a static, preset defense threshold mechanism, these highly valuable and legitimate parameter increments reflecting real extreme road conditions are easily misjudged as malicious attacks and removed by the system due to low model quality scores. This causes the global benchmark model to lose its ability to learn and generalize to extreme track environments, leading to widespread subsequent inaccuracies in judgments. To address this deep-seated technical problem, this embodiment provides a complete solution.

[0111] Specifically, for anomalous parameter packets whose model quality scores are below a preset defense threshold, before performing the filtering operation, the system also includes an adaptive aggregation calibration step based on spatiotemporal population parameter drift characteristics. The system does not immediately discard model parameter packets initially marked as anomalous on the central server side, but instead triggers a secondary diagnostic evaluation mechanism. Specifically:

[0112] The system first obtains the parameter drift feature vector of the corresponding node based on the gradient calculation of the current parameter update. At the actual algorithm execution level, the parameters of the local personalized model include the weight matrices and bias terms of multiple hidden layers in the deep neural network. To overcome the excessive computational consumption caused by directly performing nonlinear algebraic operations in the ciphertext domain, and to resolve the underlying cryptographic logic conflict that existing ciphertext domain computations cannot efficiently perform vector modulus square root and division operations while maintaining fully homomorphic encryption, this embodiment adopts a collaborative transmission and computation architecture that separates plaintext and ciphertext.

[0113] Specifically, before performing fully homomorphic encryption locally, each of the aforementioned check-in point edge computing units updates the gradients for the weights of specific key network layers in the plaintext domain. This gradient is then transformed into a one-dimensional vector sequence through tensor flattening. Subsequently, linear dimensionality reduction techniques such as local average pooling or principal component analysis are used to generate a low-dimensional parameter drift feature signature vector that characterizes the evolution direction of the core gradient. Since this feature signature vector is extracted in the plaintext domain, the edge computing unit can perform adaptive differential privacy noise enhancement to ensure that this low-dimensional vector cannot be reverse-derived back to the original multimodal physiological or visual privacy data.

[0114] Subsequently, the edge computing unit independently uploads the low-dimensional parameter drift feature signature vector to the central server via a separate differential privacy control channel, while the high-dimensional local model parameter increments are still uploaded via a ciphertext channel formed by fully homomorphic encryption. Through this plaintext-ciphertext separation architecture, the central server can obtain a low-dimensional mathematical expression that can characterize the evolution trend of the local models of each independent federated node at extremely low computing power cost without violating the overall defense logic of fully homomorphic encryption.

[0115] Optionally, a set of associated federated nodes connected to the corresponding node in spatial topology is obtained, and the parallel parameter drift feature vector uploaded by the associated federated node set within the same time slice period is extracted. In actual race road network modeling, the race track is not a simple set of planar coordinate points in Euclidean geometry, but a directed acyclic graph with strict arrival order and geographical isolation characteristics. Therefore, being connected in spatial topology does not mean proximity in absolute straight-line distance, but rather a group of adjacent nodes on the actual track route that are physically connected to the corresponding node that triggered the anomaly within a preset road segment distance, such as two adjacent timing checkpoints. The central server accurately locates the set of associated federated nodes affected by the same sudden local weather system based on the pre-set digital elevation model and topology map of the track.

[0116] Simultaneously, the system retrieves parameter drift feature vectors submitted to the central server by these related federation nodes within the same or similar time windows. These vectors are collectively referred to as parallel parameter drift feature vectors. Since the physical impact of environmental mutations inevitably possesses spatial continuity and temporal regional synchronicity, extracting these parallel parameter drift feature vectors aims to provide a group-level cross-validation reference system for the abnormal drift of isolated nodes.

[0117] Specifically, the mean spatial cosine similarity between the parameter drift feature vector and each of the parallel parameter drift feature vectors is calculated. In a multidimensional vector space, cosine similarity can accurately measure the degree of convergence between two vectors in the direction dimension, while being insensitive to the absolute difference in the magnitude of the vectors. This mathematical property is extremely suitable for the business scenario of this embodiment: due to differences in their specific locations or hardware computing power, the absolute magnitude (magnitude) of the model parameter updates at different check-in points may differ, but as long as they are driven by the same extreme environmental factors such as a rainstorm, the gradient direction of their model parameter evolution adjustment should be highly consistent. The system uses the following formula to calculate the mean spatial cosine similarity:

[0118] ;

[0119] in, This represents the calculated mean spatial cosine similarity. This represents the total number of connected nodes included in the associated federation node set; This represents the node index number in the associated federated node set, with values ​​ranging from... arrive Positive integers; This represents the parameter drift feature vector extracted from the corresponding node in this instance; Indicates the first The parallel parameter drift feature vectors corresponding to the associated federated nodes; This represents the dot product operation between the two multidimensional feature vectors mentioned above. and These represent the corresponding eigenvectors. Norm, also known as Euclidean modulus.

[0120] Using this rigorous mathematical formula, the system can obtain the average consistency index between the currently isolated abnormal node and all its surrounding topologically adjacent nodes in the direction of model evolution. If the index value is high, it proves that the current parameter drift exhibits significant regional group consensus characteristics; conversely, if the index is extremely low or even negative, it indicates that the node's parameter update direction is opposite to that of the surrounding normal nodes, which is highly likely to be due to an isolated hardware failure or malicious local area network hijacking and parameter tampering attack.

[0121] It is also important to note that when the mean spatial cosine similarity is greater than a preset group similarity bias, the corresponding node is determined to be in a local environmental mutation region. In this step, the preset group similarity bias serves as a classification threshold for determining macro-environmental influences and micro-isolated anomalies. Its value needs to be calibrated in conjunction with historical event data, such as setting it to 0.75. When the calculated mean spatial cosine similarity exceeds this bias threshold, the system logically changes the characterization of the abnormal event from "suspected node malicious contamination" to "confirmed environmental mutation in a local race segment." This judgment logic has significant practical implications in business: it not only avoids misjudging legitimate data but also enables the system to perceive the deterioration of the physical track environment in real time, providing potential safety warnings to the event organizing committee.

[0122] For nodes identified as local environmental mutation regions, an environmental mutation compensation factor is calculated, and the preset defense threshold is dynamically decayed according to this compensation factor. In conventional anomaly filtering mechanisms, the preset defense threshold presents a static, rigid constraint. When the system confirms that a node is in a local environmental mutation region, it calculates an environmental mutation compensation factor based on the overflow degree of the mean spatial cosine similarity using an exponential or linear decay function, with a value ranging from 0 to 1.

[0123] Subsequently, the system multiplies the original preset defense threshold by the environmental mutation compensation factor. This attenuation operation essentially temporarily and dynamically relaxes the anti-cheating system's tolerance boundary for drastic fluctuations in model parameters within a specific physical and spatiotemporal region. Through this mechanism, the system provides a reasonable adaptive release mechanism for data fusion under extreme environments, allowing parameter packages whose scores drop due to harsh environments to successfully cross the lowered threshold.

[0124] For example, the single-node weight factor is based on the mean of the spatial cosine similarity. Perform adaptive calibration and generate calibration weight factors. In the aggregation phase of federated learning, the influence of each node's parameters on the global model is determined by its assigned weight factors. For nodes located in regions of sudden local environmental changes, their parameter increments should not only not be discarded, but should be considered as extremely scarce and valuable extreme scenario feature samples. To enable the global baseline model to quickly absorb and adapt to such extreme environmental features and improve the model's robustness under complex conditions, the system needs to positively amplify and compensate the aggregation weights of that node.

[0125] To avoid weight overflow that may occur due to linear compensation, the system introduces a nonlinear activation function for constraint smoothing. The specific calibration calculation formula is as follows:

[0126] ;

[0127] in, This represents the calibration weight factor generated after adaptive calibration; This indicates the first one defined in the above embodiment. The single-node weight factor of each independent federated node, that is, the basic weight determined based on the amount of data and the model quality score in the initial calculation; This represents the preset local environmental change gain coefficient, used to control the upper limit of the overall gain amplitude of the weight amplification; This represents the mean spatial cosine similarity obtained through the above steps; This indicates the preset group similarity bias; This represents the hyperbolic tangent activation function.

[0128] In the actual mathematical properties of the above formula, the hyperbolic tangent activation function Able to input difference Mapped to Within the smooth interval. Since the preconditions have already been defined... ,therefore The function outputs a positive value. As the mean similarity increases, meaning the group consensus caused by environmental mutations becomes stronger, the function's output value approaches its upper limit, thus driving the basic weighting factor. Achieving a stable and controllable amplification factor. Compared to direct linear multiplication gain, The introduction of the function effectively prevents the technical risk that the weights of some nodes will be infinitely amplified in the case of extreme similarity, thereby dominating the entire global model and causing global catastrophic forgetting, and ensures the mathematical convergence and numerical stability of the aggregation process.

[0129] It should also be noted that the anomalous parameter packet is re-evaluated according to the attenuated preset defense threshold, and the increment of the ciphertext parameters that pass the re-evaluation is adjusted using the calibration weighting factor. It is incorporated into the weighted aggregation process of the incremental federated average algorithm.

[0130] In this final operation, the central server uses a dynamically decayed new threshold to allow previously blocked abnormal parameter packets to pass through. Then, it replaces the original base weight factors with adaptively calculated and amplified calibration weight factors, re-injecting these encrypted parameter increments into the update and iteration process of the global baseline model. At this point, after weighted aggregation, the global baseline model rapidly incorporates high-value feature boundaries from mutation regions, and the distribution of neuron weights within the model automatically widens its boundaries appropriately towards extreme physiological and motor feature ranges.

[0131] like Figure 5 As shown in the attached figure, the dynamic amplification and nonlinear constraint process of federated aggregation weights is intuitively illustrated in the form of a three-dimensional curved surface. The first horizontal axis represents the mean spatial cosine similarity, in dimensionless units; the second horizontal axis represents the initial single-node weight factor, in dimensionless units; and the vertical axis represents the calculated calibration weight factor, in dimensionless units.

[0132] The adaptive weight mapping surface shown in the figure has a surface color that smoothly transitions from dark blue at the bottom to bright yellow at the top. The warmer the color area, the higher the output calibration weight factor value.

[0133] By observing the waveform transformation pattern of the three-dimensional surface in the first horizontal coordinate axis direction, it can be found that when the mean spatial cosine similarity is lower than the set group similarity bias value of 0.75, the surface height is basically the same as the initial single node weight factor and is in the lower blue value area. At this time, the system maintains a normal defense state against the abnormal parameters of independent nodes.

[0134] When the mean spatial cosine similarity exceeds 0.75 and continues to increase towards the value of 1, as the characteristics of group consensus caused by environmental mutations become more prominent, the surface shape exhibits a nonlinear saturation climbing trend that first accelerates upward and then gradually flattens out.

[0135] Taking the one-dimensional cross-sectional curve in the figure with an initial single-node weight factor of constant 0.8 as an example, when the mean spatial cosine similarity increases from 0.75 to 0.95, its corresponding calibration weight factor is smoothly amplified from 0.8 to close to 1.04, without a disorderly upward surge. This surface shape with top convergence objectively confirms that the introduction of the hyperbolic tangent activation function can effectively ensure numerical stability under extreme similarity input conditions.

[0136] This mechanism enables high-value anomalous parameters in regions of local environmental mutation to be amplified at a reasonable rate, thereby driving the global baseline model to achieve rapid knowledge evolution of extreme segment features, while avoiding the system risk of global catastrophic forgetting caused by excessive weight of local nodes.

[0137] As other participants arrived and traversed the area affected by the severe weather, the newly released global baseline model already possessed prior knowledge to accommodate abnormal physiological fluctuations in this environment. At this point, even if a participant's pace decreased significantly or their heart rate fluctuated dramatically, as long as their characteristic data conformed to the newly broadened characteristic boundary distribution of the region within the global model, the system could accurately determine that they could proceed normally. This fundamentally eliminated the systemic, large-scale misjudgment caused by sudden environmental changes.

[0138] Comparing the above embodiments with existing technologies reveals that traditional distributed anti-cheating networks for sporting events and conventional federated learning defense mechanisms typically employ binary discrete static threshold filtering logic. While this crude design can defend against basic network parameter contamination, it is prone to confusing legitimate physical data mutations caused by the environment with malicious tampering when faced with highly unstructured and non-independent real-world environmental data from outdoor sporting events. This can lead to widespread system paralysis and misjudgments in extreme weather or complex terrain, severely impacting the fair progress of the event.

[0139] This invention transcends the limitations of single-point diagnosis by introducing a consensus mechanism based on geospatial topology for the drift of group feature parameters. By extracting high-dimensional tensor similarity of feature evolution, it constructs a core mathematical indicator to distinguish between isolated attacks and environmental mutations. Furthermore, the scheme, through the design of rigorous dynamic threshold decay logic and a weight adaptive smoothing calibration algorithm based on the hyperbolic tangent activation function, not only effectively preserves high-value extreme scenario features that are incorrectly rejected by traditional mechanisms, but also, through a controllable weight amplification mechanism, reverse-drives the global benchmark model to achieve rapid knowledge evolution and boundary expansion for the characteristics of adverse race segments.

[0140] This technical solution endows the competition anti-cheating network with strong adaptive repair capabilities and high system robustness in the face of sudden changes in the natural environment. It ensures that the multimodal data collaborative judgment mechanism can maintain an objective, fair and stable evaluation benchmark under any complex physical conditions, which has significant engineering application value and practical significance.

[0141] Example 3:

[0142] In the first embodiment described above, the system generates an initial global baseline model based on macroscopic environmental parameters, such as the altitude and ambient temperature of the entire race. However, in actual 100km-level mountain trail races or races with high difficulty and complex terrain, the microscopic topography of different stages at the same macroscopic altitude often varies significantly.

[0143] For example, checkpoint A on the course might be preceded by a flat, high-traction asphalt surface, while checkpoint B might be preceded by a steep, muddy uphill section covered in gravel with extremely low friction. This microscopic difference in course terrain directly and drastically alters the biomechanical energy expenditure patterns of athletes traversing this area. If the system only sends identical initial global baseline models, generated solely based on the macroscopic environment, to the edge computing units of checkpoints A and B, it will lead to severe judgment biases at checkpoint B during the initial cold start phase of the race.

[0144] Because legitimate athletes passing through the steep and muddy section of checkpoint B will inevitably have heart rate data significantly higher than the global average threshold, and their pace will inevitably be significantly reduced. This legitimate feature shift is easily misjudged as cheating by the unified initial model. To address the technical problem that the initial global benchmark model cannot accurately represent local micro-terrain changes and is prone to causing large-scale misjudgments in the early stages of the event, this embodiment provides a micro-node-level parameter reconstruction scheme based on knowledge distillation.

[0145] Specifically, after generating and distributing the initial global baseline model for this event, the process also includes a node-level parameter reconstruction step to address micro-level track terrain differences. This process lowers the model initialization granularity from the global event level to the single-node edge level, and its complete execution flow consists of the following steps:

[0146] Step 1: Collection of micro-track geographical data and generation of quantitative characteristics of physical energy consumption resistance.

[0147] The digital elevation model data and surface friction coefficient of each independent federation node segment are obtained and fused to generate microscopic terrain resistance characteristic values ​​that characterize the local physical energy consumption intensity.

[0148] Specifically, during the event preparation phase, the system is pre-connected to a high-precision geographic information system. For each independent federated node on the track, i.e., the edge computing unit of the checkpoint, the system extracts the digital elevation model (DEM) data matrix within a preset distance range (e.g., one kilometer ahead) in front of that node. The DEM data provides continuous three-dimensional absolute elevation coordinates within the race segment. By calculating the spatial partial derivative of this coordinate matrix, the system obtains the average elevation gradient distribution vector of the local race segment. Simultaneously, the system obtains the surface friction coefficient of the race segment by analyzing the remote sensing semantic segmentation map or the actual survey data of the track. The surface friction coefficient is a key parameter characterizing the physical resistance characteristics of the road surface. The differences in friction coefficients of different materials, such as ice and snow, soft sand, and dry asphalt, directly determine the energy loss rate of each step taken by the athlete.

[0149] To transform the aforementioned multi-source geospatial data into numerical constraints that can be directly accessed by deep neural networks, the system performs fusion calculations to generate microscopic terrain resistance feature values ​​characterizing local physical energy dissipation intensity. This embodiment constructs a nonlinear terrain resistance assessment algorithm, the specific calculation formula of which is as follows:

[0150] ;

[0151] in, Characterizes the micro-terrain resistance features of the corresponding independent federation nodes; Characterizes the local average gradient parameter calculated based on the digital elevation model data; Characterizes the surface friction coefficient of the extracted local race segment; Characterizes the preset slope resistance weighting coefficient; Characterizes the preset surface friction resistance weighting coefficient; Characterizes the local topological complexity constant term of the race segment; It represents an exponential function with the natural constant as its base.

[0152] In practical physical applications, the impact of slope on human energy consumption during movement exhibits a non-linear, exponential growth trend. Therefore, the system uses the local average climbing slope parameter... An exponential function calculation was applied; while the coefficient of friction of the surface medium... The smaller the surface area (e.g., slippery mud), the greater the compensatory muscle energy expenditure required by athletes to maintain balance and speed; therefore, the two are inversely correlated at the physical level. The system uses this formula to reduce the dimensionality of multi-dimensional geographical environmental variables into a single, high-confidence scalar feature: the microscopic terrain resistance feature value. A larger feature value indicates a higher physical exertion intensity in the race section ahead of the corresponding checkpoint, and the multimodal characteristic distribution of real athletes passing through this area should exhibit a more polarized state.

[0153] like Figure 6 As shown in the figure, the first horizontal axis is the spatial horizontal coordinate, with the unit being meters and the data range extending from 0 to 1000. The second horizontal axis is the spatial vertical coordinate, with the unit being meters and the data range also extending from 0 to 1000. The vertical axis is the altitude, with the unit being meters.

[0154] The three-dimensional terrain undulations shown in the image reflect the micro-topographical changes of the actual race track within a 1000-meter length and width range, with elevation data distributed between approximately 150 and 450. The colors covering the terrain surface represent the calculated terrain resistance characteristic values ​​for different areas. The legend is a spatial mapping surface of terrain resistance, and the color distribution pattern can be seen in the color bars above. The units are dimensionless.

[0155] The dark blue and cyan areas in the image are concentrated in lowland plains with gentle elevation changes and small slopes. The corresponding terrain resistance characteristic values ​​are low, remaining in the range of 3 to 4, indicating that athletes consume less physical energy when passing through such flat sections with high surface friction coefficients.

[0156] The deep red areas in the image are concentrated on steep mountain peaks with rapidly ascending terrain and extremely steep slopes. Due to the exponentially amplified slope parameter and the extremely low surface friction coefficient, the calculated terrain resistance characteristic value in this area rapidly rises to a high value range of 7 to 8. This significant correlation between color and changes in three-dimensional spatial elevation and slope gradient objectively verifies that the system can effectively extract digital elevation models and surface friction data, reliably quantifying the differences in physical energy consumption intensity at the micro-level of the racecourse.

[0157] This quantitative indicator enables each edge node to extract consistent historical prior knowledge from the federated model library based on the specific micro-terrain resistance characteristics in front of its own checkpoint during the cold start phase of the event. This effectively avoids the system misjudging and blocking legitimate high heart rate and low pace physiological data caused by local adverse terrain.

[0158] Step 2: Feature clustering and teacher network extraction based on the historical federated model library with prior constraints.

[0159] After obtaining the quantitative resistance characteristics of each checkpoint, the system uses the micro-terrain resistance characteristic value as a priori optimization constraint, performs feature clustering retrieval in the historical event federated model library, and extracts the historical feature distribution space that matches the micro-terrain resistance characteristic value as the teacher network input.

[0160] Specifically, the historical event federated model library contains stable model parameters and anonymized high-dimensional feature manifold space data generated during the operation of a large number of different types of events. Traditional model matching often relies on character retrieval of event tags, while this embodiment uses numerical micro-terrain resistance feature values ​​as retrieval anchors in continuous space. The system executes a density-based noise applied spatial clustering algorithm (DBSCAN) in the high-dimensional vector space of the historical event federated model library. During the clustering process, the system sets the micro-terrain resistance feature value of the current node as a priori optimization constraint, filtering out historical feature distribution spaces that have been fully trained and converged under road sections with the same resistance intensity from past data.

[0161] It is important to note that the extracted historical feature distribution space, which matches the current resistance, essentially contains the true boundaries and joint distribution patterns of physiological data (such as heart rate zones and blood oxygen saturation slope) and kinematic data (such as cadence and vertical amplitude) of real humans facing such extremely energy-consuming terrain. The system uses this feature space, rich in high-value local knowledge, as a parameter benchmark and loads it into the network structure to construct a teacher network model that guides subsequent parameter fine-tuning. While this teacher network model may have weak macroscopic generalization ability, it is extremely accurate in feature identification and classification boundary delineation under specific microscopic terrain resistance conditions.

[0162] Step 3: Tensor multiplication operation and parameter bias vector generation based on knowledge distillation.

[0163] After establishing a high-precision reference network, the system uses the initial global benchmark model as the student network, employs the knowledge distillation algorithm to calculate the local adaptive feature gradient, and performs tensor multiplication operation between the local adaptive feature gradient and the micro-terrain resistance feature value to generate a parameter bias vector.

[0164] Specifically, while the initial global baseline model possesses general discriminative capabilities regarding macroscopic environments, such as overall altitude and temperature, it lacks microscopic perception of local terrain. The system sets it up as a student network model. The core logic of the knowledge distillation algorithm is not to have the student network directly copy the parameters of the teacher network, but rather to have the student network learn the probability distribution of soft targets output by the teacher network when processing the characteristics of corresponding resistance road segments. The system calculates the relative entropy (e.g., KL divergence) between the output distributions of the teacher and student networks to obtain the parameter adjustment direction required to minimize the difference between the two distributions, i.e., the local adaptation feature gradient. This gradient vector indicates how the neuron weights of each neural network layer within the initial global baseline model should be shifted in the non-convex optimization space to adapt to the current specific microscopic terrain.

[0165] Furthermore, the local adaptive feature gradient only provides the direction vector for optimization, and the magnitude of the optimization must be strictly aligned with the severity of the local terrain. Therefore, the system performs a tensor multiplication operation between the calculated local adaptive feature gradient and the previously generated scalar data, namely the microscopic terrain resistance feature values. In the underlying algebraic operations, the resistance feature values ​​act as scalar multipliers, broadcasting to every component of the gradient tensor.

[0166] In practical applications, this calculation step serves two purposes: if the micro-terrain in front of the current checkpoint is extremely challenging (with a very high micro-terrain resistance characteristic value), the tensor multiplication operation will correspondingly amplify the gradient correction step size, generating a parameter bias vector with a larger numerical bias; conversely, if the resistance characteristic value is small, indicating that the local track is relatively flat, the generated parameter bias vector will also approach a zero vector. The system thus generates a parameter bias vector that highly fits the actual physical resistance of the track.

[0167] Step 4: Forward reconstruction of node-level personalized parameters and differentiated edge deployment.

[0168] Finally, the system superimposes the parameter bias vector onto the basic knowledge weights to reconstruct and generate node-level initial personalized model parameters. The reconstructed node-level initial personalized model parameters are then distributed and overlaid on the corresponding check-in point edge computing units to complete the differentiated cold start of each node.

[0169] For example, the basic knowledge weights are the underlying tensor parameters of the initial global baseline model generated from the macroscopic environmental feature parameters. The system directly injects the parameter bias vector representing local microscopic adjustment instructions into the basic knowledge weights through matrix addition. The reconstruction calculation formula is as follows:

[0170] ;

[0171] in, The initial personalized model parameters at the node level after reconstruction are represented. The weights representing the basic knowledge are macroscopic model parameters that have not undergone local terrain calibration. The micro-terrain adaptation adjustment coefficient is used to control the limit of the intervention intensity of local knowledge distillation information on the macro knowledge framework, so as to avoid overfitting local features and causing catastrophic forgetting. The local adaptive feature gradient extraction function, which is based on the knowledge distillation algorithm, encapsulates complex algebraic logic for calculating relative entropy and backpropagation differentiation. A teacher network model that represents the distribution space of the historical features is an expert system that represents the true laws of specific harsh terrain. Characterize the student network model constructed using the initial global baseline model; The micro-terrain resistance characteristic value, which represents the corresponding independent federation node, is used as a physical intensity scaling scalar factor in the calculation.

[0172] In the above reconstructed calculation formula, the complete product term on the right side of the plus sign essentially constitutes the final parameter bias vector after scaling by terrain physical features and hyperparameter constraints. Through the precise reconstruction calculation of this formula, the originally uniform and singular macroscopic model parameters have undergone geographical differentiation based on the road network topology.

[0173] The system then distributes these reconstructed, computationally generated node-level initial personalized model parameters to the corresponding checkpoint edge computing units via an encrypted communication protocol. Thus, dozens or even hundreds of edge computing units distributed across different locations on the track already possess their own exclusive initial models perfectly suited to the physical energy consumption characteristics of the specific road conditions in front of their respective checkpoints during the cold start phase before the race begins. For example, the parameter distribution boundaries of the checkpoint model deployed on the uphill gravel section have been safely widened to reasonably accommodate the abnormally high heart rate and extremely low pace combination exhibited by real runners on that section; while the parameter boundaries of the checkpoint model deployed on the flat downhill section remain tight to effectively prevent potential substitution by vehicles.

[0174] Comparing the above embodiments with existing technologies reveals that current anti-cheating and feature recognition systems for sporting events generally employ a uniform initial model distribution strategy without differentiation during model deployment. This approach is ill-suited for highly unstructured, long-distance mountain trail races, where significant fluctuations in the microscopic physical resistance of the track can cause widespread false alarms and frequent interceptions during the initial cold start phase of the race due to physical fluctuations in legitimate physiological data. This severely disrupts the normal operation of the race and forces local models to require extremely long iteration cycles to reconverge to the correct local feature boundaries.

[0175] like Figure 7 As shown in the figure, the horizontal axis represents the number of training iterations (in seconds), and the vertical axis represents the loss function value (in dimensionless units).

[0176] The blue curve of the standard global initial model in the figure shows the training performance of the macro network without the introduction of local micro-terrain features when facing extreme road conditions. Its initial loss function value is as high as 2.5, and the curve declines relatively gently. It takes about 80 iterations to gradually converge to an error level of about 0.5. This objectively reflects that the single global model has obvious feature recognition lag when dealing with sudden changes in underlying physical resistance.

[0177] In contrast, the red curve in the figure represents the network convergence performance after incorporating microscopic terrain resistance features and extracting local adaptive feature gradients through knowledge distillation. The initial loss function value of this red curve is only 1.2, significantly lower than that of conventional models, and exhibits a rapid and steep descent waveform, quickly converging to a low-level stable range of 0.2 after only about 30 iterations.

[0178] This significant numerical difference, compared with the waveform convergence speed, objectively confirms that the operation of performing tensor multiplication on the historical characteristic distribution space of the teacher network and the physical energy consumption intensity and injecting parameter biases does indeed enable the edge computing units distributed at different locations on the track to obtain a priori discrimination boundary that highly fits the local terrain before the starting signal.

[0179] This mechanism significantly reduces the amount of false alarms and intercepted data caused by sudden feature shifts during the cold start period, effectively shortens the adaptive time of the underlying algorithm, and improves the overall robustness of multimodal data collaborative judgment under complex terrain.

[0180] This invention overcomes the limitations of macroscopic environmental variables, delving into the three-dimensional physical level of microscopic terrain digital elevation models and surface friction media. The solution precisely mathematically quantifies the physical energy consumption intensity of the race track by constructing a nonlinear formula for the characteristic value of microscopic terrain resistance. Furthermore, relying on a vast historical race federated model library, it introduces a cutting-edge knowledge distillation algorithm to construct a teacher-student network architecture, and utilizes a strict tensor multiplication constraint mechanism and adjustment coefficients to achieve targeted reconstruction and bias injection of node-level model parameters.

[0181] This scheme enables the distributed defense network at the edge to possess prior feature recognition and discrimination flexibility that highly matches the local terrain of the racecourse at the moment of cold start, significantly reducing the false start rate caused by terrain extremes. At the same time, this differentiated cold start mechanism greatly shortens the adaptive convergence time of the local personalized models of each edge node in the early stage of the race, ensuring that the multimodal anti-cheating system can achieve smooth transition and accurate timing intervention under any harsh and varied complex terrain, providing deep underlying algorithmic support for the intelligent management of ultra-long-distance endurance races.

[0182] Example 4:

[0183] like Figure 8 As shown, this embodiment of the invention also provides a competition anti-cheating and accurate timing system based on multimodal data fusion. This system, as the specific hardware and software functional carrier for implementing the above method embodiments, can be deployed in the actual operating environment of long-distance outdoor endurance events.

[0184] Specifically, the system provided in this embodiment is physically deployed based on a distributed topology architecture. Its core hardware includes a central server cluster, edge computing units at checkpoints deployed at various timing checkpoints along the track, and wearable devices worn by athletes. Based on this hardware architecture, the system is logically divided into the following core functional modules:

[0185] The network construction and identity allocation module is used to build a federated computing network with a global time base server at its core, and to assign independent federated node identities and unique blockchain identity identifiers to edge computing units at each checkpoint and wearable devices of athletes.

[0186] In practical applications, this module physically runs within a central server cluster composed of multiple highly available servers. The central server, through a distributed lock election service, elects a master control node within the physical cluster to assume the function of a global time reference server. The module integrates a decentralized identity generation engine. When edge computing units at check-in points (such as industrial-grade edge gateways) and wearable devices for athletes (such as smart fitness trackers) connect to the network, they are assigned corresponding digital identity certificates and unique blockchain identity identifiers generated based on asymmetric encryption algorithms. The time calibration of each hardware device interacts in real-time with the global time reference server via a network communication interface. The aforementioned bidirectional time synchronization protocol compensation algorithm constrains physical clock errors within a preset millisecond range, thus establishing a rigid time reference for the unified spatiotemporal alignment of multimodal data across the entire network.

[0187] The data acquisition and tensor generation module is used to acquire multimodal event data within a preset time slice. The multimodal event data includes radio frequency check-in data, athlete physiological data, and visual monitoring data. The module performs irreversible desensitization processing on the multimodal event data at the data acquisition end and splices it together after spatiotemporal alignment to generate a unified multidimensional spatiotemporal feature tensor.

[0188] The module is physically connected to a variety of heterogeneous data acquisition devices: its radio frequency data acquisition end is connected to an ultra-high frequency radio frequency identification carpet antenna and timing reader deployed on the track surface; the physiological data acquisition end relies on the photoplethysmography pulse wave sensor and triaxial accelerometer integrated inside the athlete's wearable device; and the visual monitoring data acquisition end is connected to an industrial-grade high-definition network camera installed on the side of the checkpoint passage.

[0189] In the actual operation process, the built-in microcontroller unit of the athlete's wearable device and the embedded processor in the edge gateway of the check-in point concurrently read the aforementioned sensor data within a specified preset time slice. This module uses a built-in feature reduction operator to perform desensitization feature extraction on the video frames and heart rate waveforms at the edge, eliminating original images or waveform samples involving personal biometrics to achieve irreversible desensitization. Subsequently, time-series signals at different rates are input into an alignment buffer via a data bus to complete forced spatiotemporal alignment, and then concatenated to output a unified multidimensional spatiotemporal feature tensor of fixed dimensions. Simultaneously, the module controls the built-in hardware encryption chip to calculate the secure hash value of this feature tensor, which is then uploaded to the consortium blockchain distributed ledger for tamper-proof notarization via a network layer communication component.

[0190] The edge-side training and preliminary judgment module is used to perform local personalized model incremental training on the edge computing unit at the check-in point based on the unified multidimensional spatiotemporal feature tensor, and to calculate the Mahalanobis distance between the current feature vector and the normal feature distribution, generating a preliminary edge-side cheating judgment result. When the preliminary edge-side cheating judgment result is abnormal, the corresponding feature vector is extracted as a suspicious feature vector. This module is physically deployed inside the edge computing unit at the check-in point. The edge computing unit is an industrial-grade gateway device with a neural network processing unit (NPU) hardware acceleration capability.

[0191] This module leverages the parallel matrix computation power of the NPU to input locally generated unified multidimensional spatiotemporal feature tensors into a pre-built local personalized autoencoder network in real time, executing the backpropagation algorithm to achieve incremental iterative training of the local model. During the inference and decision phases, the module calls its internal linear algebra general-purpose computation unit to read the inverse of the mean vector and covariance matrix of the local normal features in real time, and calculates the local Mahalanobis distance of the current input feature vector according to the aforementioned formula.

[0192] The module is internally equipped with a hard-coded threshold comparator that compares the calculated Mahalanobis distance value with a first preset multiple threshold and a second preset multiple threshold in real time, outputting a preliminary cheating judgment result at the edge. If the judgment result falls within a preset abnormal range, the logic shunt within the module will trigger a parameter locking mechanism, extracting the currently input multidimensional feature vector and marking it as a suspicious feature vector. If the Mahalanobis distance exceeds the high-risk threshold, the module will also issue a temporary interception command to the physical interception gate or alarm indicator device on site through the industrial output interface.

[0193] The encrypted domain federated aggregation and update module is used to encrypt and differentially privacy-process the parameter increments of the local personalized model before uploading them. The updated global baseline model is then generated and distributed through encrypted domain federated aggregation on the central server. This module has a distributed collaborative structure across devices, with its front-end unit deployed at the edge computing unit of the check-in point and its back-end unit deployed in the central server cluster.

[0194] At the edge, the encryption submodule of this module calls the homomorphic encryption algorithm library to convert the weight parameter matrix updated by the local personalized model into a ciphertext tensor; at the same time, the differential privacy circuit generates Gaussian noise with the corresponding standard deviation based on the parameter sensitivity and superimposes it into the ciphertext stream.

[0195] On the central server side, the federated aggregation engine of this module receives encrypted parameter packets uploaded by each edge node while keeping the parameters encrypted. It then reads the local verification metrics attached to the packets to calculate the model quality score for that node. Subsequently, the aggregation engine executes the aforementioned incremental federated averaging algorithm, performing a weighted average of the scalar single-node weight factors of each independent federated node with the corresponding encrypted parameter increment tensor.

[0196] During this process, the module automatically performs defense checks, filtering out abnormal parameter packets whose model quality scores are below a preset defense threshold. Finally, the updated global baseline model parameter tensor is synthesized in the encrypted domain and distributed to edge computing units at various checkpoints across the network through the network distribution channel, replacing or merging the basic knowledge weights of their local personalized models to achieve a secure closed-loop update of global knowledge.

[0197] In addition, the module also includes a differential privacy control channel, as described in Embodiment 2, in which the low-dimensional parameter drift feature signature vector is extracted by the edge computing unit in the plaintext domain and uploaded independently. This enables the central server to adaptively adjust the preset defense threshold based on the mean of spatial cosine similarity and to calibrate the weight factor using the hyperbolic tangent function, thereby possessing hardware collaborative defense capabilities to resist interference from sudden changes in the local environment.

[0198] The global collaborative verification and score generation module is used to input the suspicious feature vectors into the global benchmark model, combine the cross-node spatiotemporal cross-verification results to generate a final judgment, and synchronously update the athlete's global score. Physically, this module mainly runs in the high-performance general-purpose computing unit of the cloud central server and is connected to the system's core competition database and arbitration client large-screen terminal. This module receives suspicious feature vectors uploaded from each checkpoint via a dedicated communication channel.

[0199] First, the global judgment engine is used to calculate the global Mahalanobis distance of the suspicious feature vector relative to the normal feature distribution of the latest global benchmark model. When the global Mahalanobis distance indicates a suspected violation, the spatiotemporal correlation network layer of this module will issue a retrieval request to retrieve feature snapshots of the corresponding athlete at a preset number of adjacent check-in points before and after the historical data from the core database, as well as parallel feature data of related athletes within a preset spatial radius during the same time period, and perform multi-dimensional spatiotemporal consistency cross-comparison.

[0200] The multi-level decision-maker within the module combines global Mahalanobis distance and spatiotemporal consistency comparison results to output the final judgment conclusion based on a rigorous preset condition matrix. If the final judgment conclusion is normal, the results management engine synchronously confirms the validity of its time record; if the conclusion is pending review, the system pushes a transaction package containing multimodal desensitization features, video screenshot sequences, and trajectory comparison diagrams to the chief referee's arbitration terminal for manual review; if the conclusion confirms cheating, the athlete's global results are automatically suspended or revoked, and the official event timing screen and results announcement system are updated.

[0201] In summary, the anti-cheating and accurate timing system for sports events based on multimodal data fusion provided by this invention deeply adapts complex algorithmic flows to a hardware network architecture composed of wearable devices, edge computing gateways, and a central cloud server. The modules are tightly connected physically and have causal data flow relationships, leveraging the advantages of edge hardware in low-latency processing and irreversible desensitization of high-frequency multimodal data for on-site computing, while maintaining the macro-level control capabilities of the cloud server for encrypted domain federation and global spatiotemporal cross-verification. In actual operation, the entire system effectively overcomes the engineering contradiction between harsh outdoor communication limitations and high-precision multi-dimensional anti-cheating measures, providing a highly stable equipment architecture to support the fairness and data security management of modern sports event timing.

[0202] Example 5:

[0203] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0204] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0205] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0206] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0207] The memory 103 stores a computer program corresponding to a multimodal data fusion-based anti-cheating and accurate timing method for sports events according to the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0208] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0209] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for preventing cheating and ensuring accurate timing in sports events based on multimodal data fusion, characterized in that, include: Construct a federated computing network with a global time base server at its core, and assign independent federated node identities and unique blockchain identity identifiers to edge computing units at each checkpoint and wearable devices of athletes; Multimodal event data is collected within a preset time slice, including radio frequency check-in data, athlete physiological data, and visual monitoring data. The multimodal competition data is irreversibly desensitized at the data acquisition end, and then spliced ​​together after spatiotemporal alignment to generate a unified multidimensional spatiotemporal feature tensor. Based on the unified multidimensional spatiotemporal feature tensor, local personalized model incremental training is performed in the edge computing unit of the check-in point, and the Mahalanobis distance between the current feature vector and the normal feature distribution is calculated to generate a preliminary cheating judgment result on the edge side. When the preliminary cheating determination result on the edge side is abnormal, the corresponding feature vector is extracted as a suspicious feature vector; The parameter increments of the local personalized model are encrypted and differentially privacy processed before being uploaded. The updated global benchmark model is generated and distributed through the encrypted domain federation aggregation of the central server. The suspicious feature vectors are input into the global benchmark model, and the final judgment is generated by combining the cross-node spatiotemporal cross-validation results. The athlete's global performance is updated synchronously, specifically including: Calculate the global Mahalanobis distance between the suspected feature vector and the normal feature distribution of the global benchmark model. ,in This represents the suspicious feature vector; Trigger a cross-node spatiotemporal cross-verification mechanism to retrieve the historical feature vectors of the corresponding athlete's adjacent number of check-in points and the parallel feature vectors of related athletes within a preset spatial radius for spatiotemporal consistency comparison. like Furthermore, the spatiotemporal consistency comparison passed, and it was determined to be normal. The first global distance threshold; like If the spatiotemporal consistency comparison conditions are not fully met, a review instruction will be generated, in which... This is the second global distance threshold; like Furthermore, the conditions for spatiotemporal consistency comparison were not met, which was deemed cheating; And the first global distance threshold Strictly less than the second global distance threshold .

2. The method according to claim 1, characterized in that, The step of performing irreversible anonymization processing on the multimodal competition data at the data acquisition end, and then splicing it after spatiotemporal alignment to generate a unified multidimensional spatiotemporal feature tensor includes: For the first preset time slice, the radio frequency check-in data is converted into a first preset dimension radio frequency spatiotemporal feature tensor containing timestamps, locations, and kinematic information; A second preset dimension physiological feature tensor is extracted and generated locally on the athlete's wearable device, and a third preset dimension visual space feature tensor is extracted and generated at the edge computing unit of the check-in point. The physiological feature tensor, the visual spatial feature tensor, and the radio frequency spatiotemporal feature tensor are forcibly spatiotemporally aligned according to their corresponding timestamps and then concatenated to generate a unified multidimensional spatiotemporal feature tensor with a fourth preset dimension. The secure hash value of this tensor is then calculated and stored on the blockchain.

3. The method according to claim 1, characterized in that, The Mahalanobis distance between the current feature vector and the normal feature distribution is calculated using the following formula: ； in, Indicates the current time Local Mahalanobis distance of the feature vectors; This represents the feature vector corresponding to the unified multidimensional spatiotemporal feature tensor of the current input; This represents the normal feature mean vector output by the local personalized model; The covariance matrix representing local normal features; The specific rules for generating preliminary cheating determination results on the edge side are as follows: like If it is normal, then it is considered normal. Local standard deviation, The first preset multiple threshold; like If it is, then it is determined to be abnormal and marked as the suspicious feature vector, where The second preset multiple threshold; like If so, it is judged as high risk and a temporary interception command is generated; And the first preset multiple threshold Strictly less than the second preset multiple threshold .

4. The method according to claim 1, characterized in that, The process of encrypting and performing differential privacy processing on the parameter increments of the local personalized model before uploading includes: The parameter increments of the local personalized model are extracted at a preset time period, and the parameter increments are converted into ciphertext using a homomorphic encryption algorithm; The standard deviation of Gaussian noise is dynamically matched based on the sensitivity of the parameter increment. The Gaussian noise is then injected into the encrypted parameter increment based on an adaptive differential privacy mechanism, and a differential privacy budget is set. ,in The preset privacy and security threshold; The increment of the ciphertext parameters after adding noise is uploaded to the central server via an isolated encrypted channel.

5. The method according to claim 1 or 4, characterized in that, The process of generating and distributing an updated global baseline model through federated aggregation of encrypted domains via a central server includes: On the central server, the model quality score of the corresponding node is calculated and obtained based on the local verification metrics that are synchronously uploaded by each independent federation node with the incremental ciphertext parameters. While the central server remains undecrypted, an incremental federated averaging algorithm is used to weight and aggregate the incremental ciphertext parameters uploaded by all independent federated nodes. In the weighted aggregation process, a single-node weight factor is defined. The The data volume updated at the corresponding node and the model quality score are positively correlated. Filter out abnormal parameter packets whose model quality scores are lower than a preset defense threshold, generate updated global baseline model parameters, and encrypt and distribute them to each independent federated node for parameter fusion with the local personalized model.

6. The method according to claim 1, characterized in that, The construction of a federated computing network centered on a global time base server includes: A unique global time base server is generated within the cluster through a distributed lock election mechanism. The lock lifecycle is set to the first preset duration and a re-election is performed periodically. The network latency is corrected by a two-way time synchronization protocol compensation algorithm, which constrains the time synchronization error accuracy of all independent federation nodes to within a second preset time period.

7. The method according to claim 1, characterized in that, The federated computing network is also configured with a federated model cold start mechanism, including: Obtain the macro-environmental characteristic parameters of the current event, including altitude and ambient temperature; Based on the vertical federated transfer learning algorithm, the basic knowledge weights are matched and extracted from the historical event federated model library according to the macro-environment feature parameters to generate the initial global benchmark model for this event and distribute it.

8. The method according to claim 2, characterized in that, Before collecting the multimodal event data, a device physical layer encryption and isolation binding operation is performed, including: The athlete's wearable device, which integrates a power-off self-destruct encryption unit, is physically bound to the unique blockchain identity. The global scheduling center distributes the same-source key to the high-speed volatile static random access memory of the power-off self-destruct encryption unit; when an unauthorized read instruction is detected, the power supply to the memory is instantly cut off to destroy the same-source key.

9. A competition anti-cheating and accurate timing system based on multimodal data fusion, characterized in that, include: The network construction and identity allocation module is used to build a federated computing network with a global time base server as the core, and to assign independent federated node identities and unique blockchain identity identifiers to the edge computing units at each checkpoint and the wearable devices of athletes. The data acquisition and tensor generation module is used to acquire multimodal event data within a preset time slice. The multimodal event data includes radio frequency check-in data, athlete physiological data, and visual monitoring data. The multimodal event data is irreversibly desensitized at the data acquisition end, and after spatiotemporal alignment, it is spliced ​​to generate a unified multidimensional spatiotemporal feature tensor. The edge-side training and preliminary judgment module is used to perform local personalized model incremental training on the edge computing unit of the check-in point based on the unified multidimensional spatiotemporal feature tensor, and to calculate the Mahalanobis distance between the current feature vector and the normal feature distribution, and generate the edge-side preliminary cheating judgment result. When the preliminary cheating determination result on the edge side is abnormal, the corresponding feature vector is extracted as a suspicious feature vector; The encrypted domain federated aggregation and update module is used to encrypt and perform differential privacy processing on the parameter increments of the local personalized model before uploading it, and generate an updated global benchmark model through encrypted domain federated aggregation of the central server and distribute it. The global collaborative verification and score generation module is used to input the suspicious feature vectors into the global benchmark model, combine the cross-node spatiotemporal cross-validation results to generate a final judgment, and synchronously update the athlete's global score. Specifically, it includes: Calculate the global Mahalanobis distance between the suspected feature vector and the normal feature distribution of the global benchmark model. ,in This represents the suspicious feature vector; Trigger a cross-node spatiotemporal cross-verification mechanism to retrieve the historical feature vectors of the corresponding athlete's adjacent number of check-in points and the parallel feature vectors of related athletes within a preset spatial radius for spatiotemporal consistency comparison. like Furthermore, the spatiotemporal consistency comparison passed, and it was determined to be normal. The first global distance threshold; like If the spatiotemporal consistency comparison conditions are not fully met, a review instruction will be generated, in which... This is the second global distance threshold; like Furthermore, the conditions for spatiotemporal consistency comparison were not met, which was deemed cheating; And the first global distance threshold Strictly less than the second global distance threshold .

Images