Movable wireless night vision monitoring system

Abstract

The invention discloses a movable wireless night vision monitoring system, and relates to the technical field of movable wireless night vision monitoring systems, in particular to a movable wireless night vision monitoring system which is designed for environments such as smart parks, border channels and large industrial facilities, supports real-time monitoring and data transmission, ensures clear imaging in a low-light environment, and improves the working efficiency. Meanwhile, the data security is guaranteed through a multi-level encryption and privacy protection mechanism; the mobile design is achieved, rapid deployment and flexible maintenance are facilitated, and the system adapts to various complex scenes; the night vision function is achieved through the infrared light supplementing and low-illumination imaging technology, and the face, license plate and thermal radiation characteristics are effectively captured. The method is particularly suitable for user groups such as safety management personnel and facility maintenance personnel, the safety and the monitoring efficiency are remarkably improved, and the operation cost is reduced.

Description

Technical Field

[0001] The invention relates to the technical field of mobile wireless night vision monitoring systems, in particular to a mobile wireless night vision monitoring system. Background Art

[0002] In environments such as smart campuses, border crossings, and large industrial facilities, mobile wireless night vision surveillance systems are installed on autonomous patrol robots or portable gimbals. Using infrared fill light and low-light imaging, they continuously capture faces, license plates, and thermal radiation signatures in the dark, transmitting these images back to the monitoring center in real time via a self-organizing wireless network. Because the devices are constantly in motion, channel conditions and surrounding interference fluctuate constantly. Furthermore, surveillance images naturally carry sensitive identity and behavioral information. To truly maximize the value of nighttime patrols and threat warnings, these systems must maintain clear images while ensuring secure data flow between different nodes.

[0003] A search revealed a patent application with publication number CN117216788A, which provides a blockchain-based federated learning privacy-preserving video scene recognition method. This method, which falls within the blockchain field, involves: the task issuer constructing a trusted blockchain network, with all participating nodes registering within it. During the pre-training phase, nodes train on local data based on an initial network model; the nodes are evaluated, and an adaptive hierarchical blockchain network is constructed. This layered blockchain network allows for rapid aggregation of node parameters, accelerating the training process. During the model's inference phase, the blockchain network is reused, and a 3PC secure multi-party computing cluster is constructed using verifiable random functions. Using secret sharing technology, video scene recognition is achieved without exposing private video data or models.

[0004] However, combined with existing technology and application background: Most existing solutions rely on pushing raw video streams to the cloud for centralized identification and storage. When wireless links or edge nodes lack end-to-end encryption, differential anonymity, and federated update mechanisms, hackers can eavesdrop on wireless channels or infiltrate cloud-side interfaces to steal plaintext footage and use model inference to recover individual trajectories. This has been repeatedly exposed in past attacks. Even with distributed learning, uploaded model gradients can leak original features, and research has demonstrated the need for additional security reinforcement to prevent inference attacks. Once privacy is compromised, not only can personal movements and health information be exploited maliciously, but deployers may also violate regulations such as the European General Data Protection Regulation, resulting in compliance penalties and a crisis of public trust, thus hindering the widespread adoption of mobile wireless night vision surveillance systems in public and enterprise settings. Summary of the Invention

[0005] (1) Technical problems solved To address the shortcomings of existing technologies, the present invention provides a mobile wireless night vision surveillance system. This system features a portable design, enabling rapid deployment and flexible maintenance, adapting to a variety of complex scenarios. Its night vision function utilizes infrared fill light and low-light imaging technology to effectively capture facial features, license plates, and thermal radiation signatures. This system is particularly suitable for users such as security managers and facility maintenance personnel, significantly improving safety and monitoring efficiency while reducing operating costs. This system addresses the technical issues discussed in the background technology.

[0006] (2) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: a mobile wireless night vision monitoring system, comprising: utilizing a trusted execution environment to instantly generate a one-time session key, and encrypting low-light original image frames in real time, so that the data is in a ciphertext state upon generation; On the terminal device, the fully homomorphic encryption framework and the quantized and slimmed-down YOLONano model are used to perform target detection on the ciphertext image frame, and the encrypted target box and confidence stream are output; The leakage sensitivity index and convergence contribution index are calculated in the terminal's trusted execution environment. The differential noise coefficient is then input into the privacy-utility trade-off model to adaptively adjust the Laplace noise scale. Local gradients are then noised and quantized to generate uplink data packets. The multi-band KineticMesh channel encapsulates and transmits data packets, monitors channel quality in real time, and automatically updates the routing table to achieve self-healing forwarding, ensuring that data packets reach the monitoring center. After the cloud verifies the signature of the data packet uploaded by the terminal device, it aggregates the noisy quantized gradients of multiple terminals to update the encryption model weights, and generates policy tokens and compliance metadata, which are then sent to the terminal device.

[0007] Furthermore, the monitoring device detects the movement of the target object through the motion data acquisition device or starts the image acquisition process according to a preset time plan; when the image acquisition process is started, the terminal device uses a random number generator in a trusted execution environment to generate a 256-bit one-time session key.

[0008] Furthermore, the monitoring device generates low-light original image frames through infrared fill light and low-light imaging technology, and uses a one-time session key and a randomly generated initialization vector to perform block encryption of the low-light original image frames with an Advanced Encryption Standard 256-bit key length to generate a ciphertext image frame; After encryption is completed, the terminal device immediately destroys the one-time session key from the trusted execution environment and uses a hash-based message authentication code algorithm to generate a message authentication code for the ciphertext image frame to perform verification on the ciphertext image frame.

[0009] Furthermore, the terminal device deploys a fully homomorphic encryption framework to convert the obtained encrypted low-light original image frame, i.e., the ciphertext image frame, into a ciphertext format compatible with the fully homomorphic encryption framework; The lightweight target detection model YOLONano is selected and 8-bit integer quantization is performed on it to generate quantized model parameters; the model's activation function is approximated by Chebyshev polynomials instead of the original nonlinear activation function to adapt to the operational limitations of the fully homomorphic encryption framework.

[0010] Furthermore, the quantized YOLONano model is used to perform homomorphic convolutional feature extraction, homomorphic pooling, target positioning and classification on the converted ciphertext image frame in the fully homomorphic encryption framework, and outputs the encrypted target box and confidence stream; by adopting a block processing method for the input ciphertext image frame, combined with model quantization and parallel processing of the fully homomorphic encryption framework.

[0011] Furthermore, the terminal device calculates the local gradient using the encrypted target frame and confidence flow in a trusted execution environment, calculates the Fisher information matrix trace based on the local gradient, and performs virtual reconstruction to obtain the average peak signal-to-noise ratio of the reconstructed frame and the encrypted original frame after decryption, and calculates the leakage sensitivity index based on this.

[0012] Furthermore, the terminal device obtains the historical global gradient sent from the cloud and calculates the cosine similarity between the local gradient and the historical global gradient; the instantaneous loss reduction rate after applying the local gradient to update the model is calculated on the verification set of the terminal device, and the convergence contribution index is calculated based on this.

[0013] Furthermore, the terminal device inputs the leakage sensitivity index and convergence contribution index into the pre-trained privacy-utility balance model and outputs the differential noise coefficient. The Laplace noise scale is calculated based on the differential noise coefficient, the global sensitivity of the local gradient, and the differential privacy budget. Laplace noise is added to each component of the local gradient and quantized to generate a noisy gradient. The terminal device encapsulates the quantized noisy gradient set into an uplink data packet and adds the terminal device signature.

[0014] Furthermore, when transmitting data packets, the terminal device sends beacon frames at regular intervals to monitor signal strength, signal-to-noise ratio, and packet loss rate, and calculate channel quality indicators; When the channel quality index is lower than the preset threshold, the terminal device selects the path with the highest channel quality index or switches to the backup frequency band according to the routing table, updates the routing table and retransmits the data packet using the new path or new frequency band; After receiving the transmission data packet, the monitoring center sends a confirmation frame. If the terminal device does not receive the confirmation frame within the specified time, it resends the transmission data packet. When the upper limit of retransmission times is exceeded, an alarm is triggered.

[0015] Furthermore, the real-time data collected by the data acquisition device is acquired through the data processing unit, the real-time data is analyzed based on a preset algorithm, an analysis result is generated, and the analysis result is transmitted to the display module for real-time display; The data acquisition device continuously monitors the target parameters, the data processing unit performs data processing at fixed time intervals, and the display module synchronously updates the output of the analysis results.

[0016] (3) Beneficial effects The present invention provides a portable wireless night vision monitoring system, which has the following beneficial effects: Introducing instant key-driven encrypted imaging during the image acquisition phase ensures that low-light raw image frames are in a ciphertext state from the moment they are generated. By encrypting image data in real time on the acquisition side, the threats of wireless sniffing and local reading attacks are effectively prevented, laying a solid security foundation for subsequent processing and transmission.

[0017] Using homomorphic reasoning for encrypted object detection, the system leverages the quantized and streamlined YOLONano model and a fully homomorphic encryption framework on the terminal device to detect objects in encrypted image frames, outputting encrypted object bounding boxes and confidence streams. Through model optimization and parallel processing, the system achieves efficient object detection while maintaining real-time performance.

[0018] By calculating the leakage sensitivity index and convergence contribution index and inputting them into the privacy-utility balance model to adaptively adjust the Laplace noise scale, the accuracy and convergence speed of the model are maintained, providing high-quality input data for cloud aggregation.

[0019] Multi-band channels and real-time channel quality monitoring ensure reliable uplink data packet transmission in mobile and obstructed scenarios. When network conditions change, the system automatically updates routing tables and switches paths or frequency bands to maintain real-time and stable data transmission, making it particularly suitable for the complex mobile environments common during night patrols.

[0020] By verifying the signature of the data packet uploaded by the terminal device, aggregating the noisy quantization gradients of multiple terminals, updating the encryption model weights, generating policy tokens and compliance metadata, and securely sending the updated model to the terminal device, the system can simultaneously achieve the triple goals of high-definition imaging, online learning, and trustworthiness in scenarios such as smart campuses and night patrols at border crossings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a structural diagram of the mobile wireless night vision monitoring system of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] See also Figure 1 The present invention provides a portable wireless night vision monitoring system, comprising: Step 1. The monitoring device detects the movement of the target object through a motion sensor or starts the image acquisition process according to a preset time plan; when the image acquisition process is started, the terminal device uses a random number generator to generate a 256-bit one-time session key in a trusted execution environment; the monitoring device generates low-light original image frames through infrared fill light and low-light imaging technology, and uses the one-time session key and a randomly generated initialization vector to perform block encryption of the low-light original image frames with an Advanced Encryption Standard 256-bit key length to generate ciphertext image frames; after the encryption is completed, the terminal device immediately destroys the one-time session key from the trusted execution environment; the system uses a hash-based message authentication code algorithm to generate a message authentication code for the ciphertext image frame to verify the integrity and authenticity of the ciphertext image frame.

[0024] The step 1 includes the following: Step 101: Shooting action trigger mechanism During operation, the monitoring device detects the movement of the target object through the built-in motion sensor, or starts the encrypted imaging process of image acquisition according to a pre-set time plan.

[0025] Specifically, when a motion sensor detects movement in the environment, it generates a hardware interrupt signal, which is transmitted directly to the monitoring device's processor, prompting the system to immediately enter image acquisition. Similarly, if a timed schedule is used, the device's internal clock generates a trigger signal at a predetermined time, initiating the same acquisition process. This hardware interrupt signal generation is based on the processor's real-time monitoring of sensor inputs, ensuring that the time from signal generation to system response is kept to milliseconds, meeting the requirements for rapid target capture during nighttime patrols.

[0026] Using hardware interrupt signals as a trigger mechanism significantly shortens the time between detection and response, ensuring timely acquisition of image data in dynamic scenes with rapidly moving targets and avoiding information loss due to latency. This fast-response trigger mechanism allows surveillance equipment to effectively capture targets within the real-time requirements of nighttime patrols, improving system reliability and environmental adaptability.

[0027] Step 102: Generate a one-time session key A trusted execution environment (TEE) is set up within the monitoring device's terminal to generate one-time session keys. This environment is an independent, hardware-isolated area equipped with a dedicated random number generator. When the image acquisition process initiates, the TEE uses the random number generator to generate a 256-bit random key as the one-time session key. This key complies with the Advanced Encryption Standard (AES) standard for symmetric encryption algorithms, specifically a 256-bit key length version. Each time an image acquisition action is triggered, the system generates a new key with no mathematical or logical connection to the previous key, ensuring the independence of each key. Hardware isolation within the TEE prevents interference with the key generation process from external operating systems or malware.

[0028] When used, the independence of the one-time session key and the security of the generation process effectively reduce the possibility of the key being cracked or leaked, thereby improving the confidentiality of image data.

[0029] Step 103: Encrypting low-light original image frames The surveillance device uses infrared fill light technology and low-light imaging technology to generate a frame of raw image data, called a low-light raw image frame. After generation, this image frame is immediately encrypted in the trusted execution environment.

[0030] The specific process is as follows: Using the previously generated one-time session key, the low-light original image frame is encrypted to generate the corresponding ciphertext image frame. Encryption utilizes the Advanced Encryption Standard (AESS) block cipher mode with a 256-bit key length, specifically the cipher block chaining mode. In this mode, the system first generates a random initialization vector (IV), combines it with the key, and encrypts the image frame data segment by segment, ensuring that each encryption result is unique. After encryption is complete, the ciphertext image frame is stored in a temporary memory buffer on the terminal device for subsequent processing or transmission.

[0031] Encrypting image data immediately after it is generated ensures that the data always exists in ciphertext form within the terminal device, preventing it from being exposed to potential malicious visitors in an unencrypted state. This allows encryption protection to be implemented from the source of the data, ensuring that low-light original image frames are secure at all stages after generation, preventing unauthorized reading or theft.

[0032] Step 104: Key destruction and management After the one-time session key completes the encryption of the low-light original image frame, it is immediately deleted from the trusted execution environment and is not retained. Specifically, after confirming that the ciphertext image frame has been generated and stored in a temporary buffer, the system automatically clears all data related to the key in the memory. If it is subsequently necessary to decrypt the ciphertext image frame at the monitoring center, the one-time session key and initialization vector are sent separately to the authorized receiver in an encrypted manner through the Transport Layer Security Protocol version 1.3. Alternatively, the terminal device and the monitoring center negotiate through the Diffie-Hellman key exchange protocol to generate a shared key for decrypting the ciphertext image frame. The key transmission or negotiation process is completed in an encrypted channel to ensure security.

[0033] Destroying the key immediately after encryption reduces the time it remains in the system, thereby reducing the risk of unauthorized access. Using secure protocols to transmit keys or negotiate shared keys ensures the authorization of the decryption process. Key destruction reduces the possibility of leakage, and secure transmission and negotiation mechanisms ensure that only authorized recipients can access image data, enhancing the overall security of the system.

[0034] Step 105: Data integrity and authenticity protection While generating the ciphertext image frame, the system calculates a message authentication code (MAC) to verify the integrity and authenticity of the data. This method uses a hash-based message authentication code algorithm, specifically the 256-bit Secure Hash Algorithm (SHA-2), with a one-time session key as input. The ciphertext image frame is hashed to generate a fixed-length authentication code. This authentication code is stored along with the ciphertext image frame in a temporary buffer or sent to the receiver during transmission. After receiving the data, the receiver repeats the hash operation and compares the authentication code to confirm that the data has not been tampered with or forged during transmission.

[0035] When used, the calculation of a message authentication code provides an additional layer of verification for the encrypted image frame, ensuring that the data remains in its original state during transmission or storage and is not maliciously modified. The application of a message authentication code enhances the credibility of the encrypted image frame, enabling the surveillance system to resist attacks such as data tampering or forgery, ensuring the reliable use of data.

[0036] Image acquisition is triggered by motion sensors or a scheduled schedule. A one-time session key is generated within a trusted execution environment (TEE). Low-light raw image frames are encrypted in real time using a 256-bit key length version of the Advanced Encryption Standard (AES) to generate ciphertext frames. After encryption, the key is immediately destroyed, and the data integrity and authenticity are verified using a message authentication code. This process ensures that image data is encrypted from the moment it is generated and is tamper-resistant. This effectively mitigates the risk of unauthorized access or data interception during nighttime patrols, providing secure and efficient data processing capabilities for the surveillance system.

[0037] Step 2: The terminal device deploys a fully homomorphic encryption framework to convert the encrypted low-light original image frame obtained from the previous step, i.e., the ciphertext image frame, into a ciphertext format compatible with the fully homomorphic encryption framework; the lightweight target detection model YOLONano is selected and quantized to 8-bit integers to generate quantized model parameters; the activation function of the model is approximated by Chebyshev polynomials to replace the original nonlinear activation function to adapt to the computational limitations of the fully homomorphic encryption framework; the quantized YOLONano model is used to perform homomorphic convolution feature extraction, homomorphic pooling, target positioning and classification on the converted ciphertext image frame in the fully homomorphic encryption framework, and outputs the encrypted target box and confidence stream; the real-time performance of target detection is ensured by adopting a block processing method for the input ciphertext image frame and combining the parallel processing of model quantization and the fully homomorphic encryption framework.

[0038] The second step includes the following: Step 201: Input the encrypted image frame The monitoring device obtains the encrypted low-light original image frame from the previous step. This encrypted low-light original image frame is called the ciphertext image frame. The ciphertext image frame is encrypted using a one-time session key generated in the terminal's trusted execution environment and a randomly generated initialization vector, using the Advanced Encryption Standard block cipher mode (specifically, cipher block chaining mode) with a 256-bit key length. The ciphertext image frame serves as input for subsequent target detection, ensuring that the data remains encrypted during processing and preventing the exposure of sensitive information on the terminal device.

[0039] When used, the ciphertext image frame is used as input, and subsequent processing can be performed without decrypting the data, thereby protecting sensitive information in the image from being obtained by unauthorized access; by keeping the data encrypted, target detection tasks can be securely performed on resource-constrained mobile devices.

[0040] Step 202: Initialize the fully homomorphic encryption framework Deploy a fully homomorphic encryption framework, such as SEAL or HElib, on the terminal device to support mathematical operations on ciphertext. Fully homomorphic encryption technology allows operations such as addition and multiplication to be performed directly on encrypted data without first decrypting the data, thereby ensuring data privacy. First, the ciphertext image frame is converted into a ciphertext format compatible with the fully homomorphic encryption framework. The specific process includes mapping the bit stream of the ciphertext image frame to a fully homomorphic encrypted plaintext polynomial ring and re-encrypting it using the fully homomorphic encrypted public key to ensure that subsequent target detection operations can be performed in the ciphertext domain.

[0041] The introduction of a fully homomorphic encryption framework enables complex mathematical operations to be performed on ciphertext image frames without exposing the plaintext data, ensuring data processing security. By performing operations on ciphertext, the system avoids the need to decrypt the image frames on the terminal device, significantly reducing the risk of data leakage while providing security for subsequent target detection.

[0042] Step 203: Deployment of the Quantified Slimming YOLONano Model The lightweight object detection model YOLONano was selected and quantized to 8-bit integers to generate quantized model parameters, reducing the model's computational complexity and storage requirements. This quantized model significantly reduces model size and inference time while maintaining high detection accuracy. Furthermore, given that the fully homomorphic encryption framework only supports addition and multiplication operations, the system adjusts the model's activation function, replacing the original nonlinear activation function (such as ReLU) with a Chebyshev polynomial approximation to ensure correct model execution in the ciphertext domain.

[0043] Model quantization can effectively reduce computing resource consumption, making it suitable for resource-constrained terminal devices; the approximate adjustment of the activation function is to adapt to the computational limitations of the fully homomorphic encryption framework; when the quantized YOLONano model runs on the terminal device, it can complete the target detection task with lower computing overhead and faster speed, while ensuring the compatibility of the model in the ciphertext domain through the adjustment of the activation function.

[0044] Step 204: Calculation of ciphertext domain target detection Within the fully homomorphic encryption framework, the quantized YOLONano model is used to perform object detection on the converted ciphertext image frames. The specific computational process includes the following steps: First, a homomorphic convolution operation is performed on the ciphertext, using the quantized convolution kernel to extract the image's feature information; second, a homomorphic max pooling operation is applied to the feature map after the convolution operation to reduce the feature dimension and retain key information; finally, the encrypted target box coordinates and classification confidence are calculated using a homomorphic fully connected layer and a polynomial approximation activation function. All of these operations are completed in the ciphertext domain, without the need to decrypt the image frame, and the encrypted target detection results, including the encrypted target box and confidence stream, are ultimately output.

[0045] When in use, target detection is performed in the ciphertext domain, ensuring that the data is always encrypted during the entire processing process, avoiding the risk of plaintext exposure; through homomorphic operations, the system can complete complex target detection tasks while protecting data privacy, meeting the dual needs of security and real-time performance in night patrol scenarios.

[0046] Step 205: Format and output the encryption result The output of ciphertext-domain object detection, namely the encrypted target bounding box and confidence stream, is converted into a ciphertext format suitable for transmission. This formatting process ensures that the detection results can be directly used for subsequent differential noise gradient encoding or securely transmitted to the cloud without the need for decryption on the terminal device. The formatted encrypted detection results contain the encrypted target bounding box coordinates and classification confidence, maintaining data confidentiality.

[0047] When in use, the test results are kept in an encrypted state and formatted, which helps to continue to protect the privacy of the data during subsequent processing or transmission; the output of encrypted test results not only provides secure data input for the next step, but also ensures the consistent protection of data throughout the entire system process, enhancing the overall security of the system.

[0048] In order to ensure the real-time performance of the target detection process, a variety of optimization measures are taken.

[0049] First, the amount of computation is reduced through model quantization technology, allowing the model to run quickly on terminal devices; second, combined with the parallel processing capabilities of the fully homomorphic encryption framework, the system can process multiple data blocks simultaneously, shortening the overall processing time; in addition, the system uses a block processing method for the input ciphertext image frame, dividing the image frame into multiple small blocks, and performing homomorphic operations on each block, further reducing the delay of a single operation and ensuring that the detection results can be output in a timely manner in the dynamic scene of night patrols.

[0050] Real-time performance is a core requirement of night patrol monitoring systems. The introduction of optimization measures aims to ensure that the system can quickly complete target detection on resource-constrained terminal devices. Through computational optimization and parallel processing, the system can meet real-time requirements while ensuring data security, ensuring that monitoring equipment can respond to and process target information in a dynamic environment in a timely manner.

[0051] In use, the system deploys a fully homomorphic encryption framework and a quantized and streamlined YOLONano model on the terminal device to perform target detection on encrypted low-light raw image frames, generating encrypted target bounding boxes and confidence streams. This entire process is completed within the ciphertext domain, eliminating the need to decrypt image frames to plaintext and ensuring the security of sensitive data in scenarios such as smart campuses or nighttime border patrols. Furthermore, through optimization measures such as model quantization and parallel processing, the system achieves real-time target detection on resource-constrained terminal devices, providing efficient and secure data processing capabilities for surveillance systems.

[0052] Step 3: The terminal device calculates the local gradient using the encrypted target frame and confidence flow in the trusted execution environment. Based on the local gradient, it calculates the Fisher information matrix trace and performs virtual reconstruction to obtain the average peak signal-to-noise ratio of the reconstructed frame and the encrypted original frame after decryption. The leakage sensitivity index is calculated based on this. The terminal device obtains the historical global gradient sent from the cloud, calculates the cosine similarity between the local gradient and the historical global gradient, and calculates the instantaneous loss reduction rate after applying the local gradient to update the model on the terminal device's validation set. Based on this, the convergence contribution index is calculated. The terminal device inputs the leakage sensitivity index and convergence contribution index into the pre-trained privacy-utility trade-off model, outputs the differential noise coefficient, and calculates the Laplace noise scale based on the differential noise coefficient, the global sensitivity of the local gradient, and the differential privacy budget. Laplace noise is added to each component of the local gradient and quantized to generate a noisy gradient. The terminal device encapsulates the quantized noisy gradient set into an uplink data packet and adds the terminal device signature.

[0053] The step three includes the following: Step 301: Prepare input data The monitoring device obtains the encrypted target bounding box and confidence stream from the previous step as input for the terminal device to calculate the gradient. The terminal device uses the encrypted target bounding box and confidence stream along with the current model parameters to calculate the local gradient. The local gradient represents the direction of the model parameter update on the current batch of encrypted data and is the target of the subsequent noise processing.

[0054] Local gradients are key information for model learning. By calculating gradients on encrypted data, continuous model learning can be supported without exposing the original data. Using encrypted target boxes and confidence streams as input ensures that data remains encrypted during processing, preventing sensitive information from being exposed on terminal devices.

[0055] Step 302: Calculate the leakage sensitivity index In the trusted execution environment of the terminal device, the Fisher information matrix trace is calculated for the local gradient. The Fisher information matrix measures the sensitivity of the gradient to changes in model parameters. Its trace is the sum of the diagonal elements of the matrix, which represents the total sensitivity. The system performs a virtual reconstruction, simulating an attacker using local gradients to restore the original frame, and calculates the average peak signal-to-noise ratio of the reconstructed frame and the encrypted original frame after decryption. The peak signal-to-noise ratio reflects the quality of reconstruction, and the lower the value, the higher the risk of privacy leakage. The leakage sensitivity index is calculated by multiplying the ratio of the Fisher information matrix trace to the preset maximum value by the normalized difference between the peak signal-to-noise ratio and the preset maximum value. The value range is 0 to 1, and the larger the value, the higher the risk of privacy leakage.

[0056] Through the Fisher information matrix trace and the peak signal-to-noise ratio of virtual reconstruction, the privacy leakage risk in the local gradient can be quantified, providing a basis for subsequent noise adjustment; the calculation of the leakage sensitivity index provides a quantitative indicator for privacy protection, ensuring that the system can adjust the noise intensity according to the actual risk while protecting privacy.

[0057] Step 303: Calculate the convergence contribution index The terminal device obtains the historical global gradient sent by the cloud after the previous round of aggregation and calculates the cosine similarity between the local gradient and the historical global gradient. The cosine similarity is obtained by dividing the dot product of the local gradient and the historical global gradient by the product of their L2 norms. It ranges from -1 to 1. The closer the value is to 1, the higher the directional consistency. On the validation set of the terminal device, the instantaneous loss reduction rate after applying the local gradient to update the model is calculated, that is, the difference between the loss value before the update and the loss value after the update divided by the loss value before the update. The convergence contribution index is obtained by normalizing the cosine similarity to 0 to 1 and multiplying it by the instantaneous loss reduction rate. The value range is 0 to 1. The larger the value, the higher the contribution to model convergence.

[0058] Cosine similarity and the instantaneous loss reduction rate can be used to assess the contribution of local gradients to model convergence, ensuring that the model's learning effect is maintained while protecting privacy. Calculating the convergence contribution index provides a utility metric for the system, ensuring that during noise addition, the noise intensity can be adjusted based on the actual contribution of the gradient.

[0059] Step 304: Privacy-Utility Balance Model The leakage sensitivity index and convergence contribution index are input into a pre-trained privacy-utility trade-off model, which outputs the differential noise adjustment coefficient. The privacy-utility trade-off model is a neural network trained using historical data to increase noise when the leakage sensitivity index is high and reduce noise when the convergence contribution index is high, thereby balancing privacy and utility. The differential noise adjustment coefficient is a positive real number; larger values ​​indicate greater noise.

[0060] The privacy-utility trade-off model adaptively adjusts noise intensity based on the leakage sensitivity index and convergence contribution index, ensuring privacy protection while maximizing the model's learning utility. By dynamically adjusting noise intensity using a pre-trained neural network, it ensures a balance between privacy protection and model learning effectiveness in different scenarios.

[0061] Step 305: Add Laplace noise The Laplace noise scale is calculated based on the differential noise coefficient. The Laplace noise scale is calculated by dividing the global sensitivity of the local gradient by the product of the differential privacy budget and the differential noise coefficient. The global sensitivity is the maximum change in the local gradient determined in advance through analysis, and the differential privacy budget controls the strength of privacy protection. A Laplace distributed random variable with mean zero and scale equal to the Laplace noise scale is added to each component of the local gradient to generate a noisy gradient. The noisy gradient components are quantized to reduce the amount of transmitted data.

[0062] When used, the addition of Laplace noise provides differential privacy protection for local gradients, ensuring that sensitive information in the original data is not leaked when the gradients are uploaded. By adaptively adjusting the noise scale, the safety and effectiveness of night patrols in scenarios such as smart campuses and border crossings can be ensured.

[0063] Step 306: Uplink data packet encapsulation The quantized, noisy gradient set is encapsulated as an uplink data packet, and the terminal device signature is added to the uplink data packet to ensure data integrity and source authenticity. The uplink data packet serves as the output of step S3 and is then transmitted to the cloud via a self-healing multi-hop mesh link. Through encapsulation and signing, the system ensures the security and authenticity of the uplink data packet during transmission, preventing data tampering or forgery. The encapsulation and signing of the uplink data packet provide reliable input for cloud-based aggregation, ensuring the security and effectiveness of model updates.

[0064] When in use, the encrypted target box and confidence flow are used to calculate the local gradient. The leakage sensitivity index is calculated by the Fisher information matrix trace and the peak signal-to-noise ratio of the virtual reconstruction. The convergence contribution index is calculated by the cosine similarity and the instantaneous loss reduction rate. Both are input into the privacy-utility balance model to obtain the differential noise coefficient. The Laplace noise scale is set accordingly to add noise to the local gradient and quantize it, and finally an uplink data packet is generated. This process provides privacy-protected gradient information for night patrol scenarios in smart parks and border passages, and also supports cloud-based model updates.

[0065] Step 4. When transmitting a data packet, the terminal device sends a beacon frame at fixed time intervals to monitor the signal strength, signal-to-noise ratio, and packet loss rate, and calculates the channel quality index. When the channel quality index is lower than the preset threshold, the terminal device selects the path with the highest channel quality index according to the routing table or switches to the backup frequency band. After updating the routing table, the terminal device retransmits the data packet using the new path or frequency band. After receiving the data packet, the monitoring center sends a confirmation frame. If the terminal device does not receive the confirmation frame within the specified time, it resends the data packet. When the upper limit of the retransmission times is exceeded, an alarm is triggered.

[0066] The step 4 includes the following contents: Step 401: Data packet encapsulation The monitoring device obtains the uplink data packet from the previous step. The uplink data packet contains the noisy quantized gradient set and the terminal device signature. The uplink data packet is combined with the encrypted alarm stream (for example, encrypted data of abnormal behavior alarms) and encapsulated into a transmission data packet. The transmission data packet is encapsulated using the multi-band KineticMesh protocol format, ensuring that the data packet can seamlessly switch between multiple frequency bands such as 2.4 GHz, 5 GHz, and sub-GHz to adapt to different channel conditions. The encapsulation process combines the uplink data packet and the encrypted alarm stream into a single data unit and adds protocol header information to identify the source, destination, and frequency band support characteristics of the data packet.

[0067] Combining and encapsulating uplink data packets with encrypted alarm streams can reduce the network resource usage caused by multiple independent transmissions. At the same time, the multi-band KineticMesh protocol format supports flexible switching between different frequency bands, improving transmission adaptability. Through combined encapsulation and multi-band support, the system can maintain data transmission stability in complex channel environments, ensuring that monitoring equipment can continue to transmit data in the movement and obstruction scenarios of night patrols.

[0068] Step 402: Multi-band KineticMesh channel establishment A multi-band KineticMesh network is established between end devices and the monitoring center, supporting 2.4 GHz, 5 GHz, and sub-GHz frequency bands. Nodes in the mesh network (including end devices, relay nodes, and the monitoring center) automatically discover neighboring nodes using the KineticMesh protocol and dynamically maintain a routing table. This routing table records all available paths from end devices to the monitoring center, annotating each path with the frequency band used and the corresponding channel quality indicator. Automatic discovery is based on periodic broadcast messages between nodes, while dynamic maintenance is achieved through real-time updates of path information.

[0069] When in use, the establishment of a multi-band KineticMesh mesh network can achieve flexible selection between different frequency bands and paths, enhancing the network's anti-interference ability; by dynamically maintaining the routing table, the transmission path can be quickly adjusted when the channel conditions change, ensuring the reliability of data transmission and adapting to the movement and obstruction during night patrols in smart parks and border passages.

[0070] Step 403: Real-time monitoring of channel quality When transmitting data packets, the terminal device monitors the channel quality of the current frequency band in real time. The evaluation parameters include signal strength, signal-to-noise ratio and packet loss rate.

[0071] The monitoring process involves sending beacon frames at fixed intervals (e.g., 1 second). The receiver then reports back measured signal strength and signal-to-noise ratio (SNR). The receiver then collects statistics on the transmission of several recent data packets and calculates the packet loss rate. The packet loss rate is the number of unsuccessfully received packets divided by the total number of transmitted packets. The channel quality index (CQI) is calculated by normalizing the signal strength measurement to a range of 0 to 1, normalizing the SNR measurement to a range of 0 to 1, subtracting the packet loss rate from 1, and averaging the three to obtain the CQI. The CQI ranges from 0 to 1, with lower values ​​indicating poorer channel quality.

[0072] During use, by real-time monitoring of channel quality, channel attenuation or interference problems can be discovered in a timely manner, providing a decision-making basis for subsequent path adjustments; by comprehensively evaluating signal strength, signal-to-noise ratio and packet loss rate, the transmission performance of the current frequency band can be accurately judged, ensuring that countermeasures are taken when channel quality deteriorates to maintain the continuity of data transmission.

[0073] Step 404: Self-healing forwarding mechanism A pre-set channel quality threshold (for example, 0.6) triggers the self-healing forwarding mechanism when the channel quality index falls below this threshold. The terminal device selects an alternative path or switches frequency bands based on the routing table, prioritizing the path with the highest channel quality index in the routing table. If no path with a higher channel quality index exists in the routing table, the device switches to an alternative frequency band (for example, from 2.4 GHz to 5 GHz) and reestablishes the connection. The routing table is updated after the path or frequency band switch, and the terminal device retransmits the data packet using the new path or frequency band.

[0074] When in use, the self-healing forwarding mechanism automatically adjusts the transmission path or frequency band to maintain data transmission reliability even when channel quality degrades. This self-healing forwarding system maintains data connectivity despite movement and obstruction during nighttime patrols, ensuring that transmitted data packets reliably reach the monitoring center and meet real-time requirements.

[0075] Step 405: Data packet reception and confirmation After receiving a transmission data packet, the monitoring center sends an acknowledgment frame to the terminal device, notifying it of successful data packet reception. If the terminal device does not receive the acknowledgment frame within a specified time (for example, 2 seconds), it resends the transmission data packet and records the number of retransmissions. When the number of retransmissions exceeds the maximum (for example, 5), the system triggers an alarm, notifying operations personnel to check the network status. The retransmission process reuses the current path and frequency band until a acknowledgment frame is received or the retransmission limit is reached.

[0076] When in use, through the confirmation frame and retransmission mechanism, the system can ensure that the transmitted data packet is successfully received by the monitoring center to prevent data loss. The application of the confirmation frame and retransmission mechanism enhances the reliability of data transmission and ensures that the transmitted data packet can still reach the monitoring center when the channel conditions are poor.

[0077] Step 406: Link status record The terminal device records the channel quality indicators, selected path, and number of retransmissions for each transmission, storing them in a local log. This local log stores transmission-related information in a time series format and periodically uploads it to the cloud server for subsequent network optimization and troubleshooting. This logging process is performed automatically by the terminal device, and the upload cycle can be set based on actual needs (for example, once an hour).

[0078] Recording link status information provides data support for network performance optimization and troubleshooting, improving system maintainability. By recording link status, the system can analyze long-term network performance, improve transmission strategies, and maintain efficient data transmission capabilities.

[0079] By encapsulating uplink data packets and encrypted alarm streams into transmission data packets, and utilizing multi-band KineticMesh mesh network channels and real-time calculated channel quality indicators, the routing table is automatically updated and the path or frequency band is switched when channel attenuation or packet loss is detected. This ensures that the transmission data packets can reliably reach the monitoring center in the movement and obstruction scenarios of night patrols in smart campuses and border passages, ensuring the real-time and security of transmission.

[0080] Step 5: The processor acquires the real-time data collected by the sensor, analyzes the real-time data based on a preset algorithm, generates analysis results, and transmits the analysis results to the display module for real-time display. The sensor continuously monitors the target parameters, the processor performs data processing at fixed time intervals, and the display module synchronously updates the output of the analysis results. Among them, the names uniformly use "sensor" to refer to the data acquisition device, "processor" to refer to the data processing unit, "display module" to refer to the result display unit, "real-time data" to refer to the immediate information collected by the sensor, and "analysis results" to refer to the output data generated by the processor based on the algorithm; The step five includes the following: Step 501: Gradient package signature verification The cloud receives uplink data packets from multiple end devices. Each uplink data packet contains a noisy set of quantized gradients and the end device's digital signature. The cloud verifies the digital signature in the uplink data packet using a public key pre-shared with each end device to confirm the source authenticity and data integrity of the uplink data packet. Uplink data packets that have been verified to be authentic and untampered with are accepted and enter the subsequent aggregation process. Uplink data packets that fail verification are discarded, and the cloud logs the discard event as an exception.

[0081] By verifying the digital signature, we can ensure that the uplink data packet is indeed sent by a legitimate terminal device and has not been maliciously tampered with during transmission, thereby preventing attackers from forging or modifying gradient data and ensuring the credibility of the gradient data received by the cloud.

[0082] Step 502: Multi-terminal gradient aggregation The cloud aggregates the signature-verified noisy quantized gradients to generate a global gradient.

[0083] The aggregation process uses a federated averaging algorithm. Specifically, the noisy quantized gradient sets uploaded by all participating devices are summed up one by one. This sum is then divided by the number of participating devices to produce the global gradient. The entire aggregation process is performed in a secure cloud computing environment, ensuring that the noisy quantized gradient sets and the resulting global gradients are not accessible to unauthorized entities. The resulting global gradients more comprehensively represent the training results of all devices, improving the overall performance and adaptability of the model while protecting the privacy of the devices.

[0084] Step 503: Update encryption model weights The cloud uses the generated global gradient to update the stored encrypted model weights. The update process is as follows: Subtract the product of the learning rate and the global gradient from the currently stored encrypted model weight to obtain the updated encrypted model weight; the learning rate is a fixed value pre-set by the cloud based on the model training strategy, which is used to control the update amplitude; after the update is completed, the updated encrypted model weight remains encrypted and is protected using the key agreed upon with the terminal device.

[0085] When in use, the update of model weights is completed in an encrypted state, which can ensure that the model parameters will not be accessed by unauthorized people when processed and stored in the cloud, thereby maintaining the confidentiality of the data.

[0086] Step 504: Policy token generation The cloud generates a policy token for each terminal device. This token contains model update permissions and access control information. This token is digitally signed to ensure it cannot be forged or tampered with. It includes information such as the terminal device ID, model version number, and expiration date. Each policy token is bound to the corresponding terminal device's public key for identity and permission verification during subsequent issuance.

[0087] When in use, a clear permission management mechanism is provided for terminal devices by generating policy tokens to ensure that only authorized terminal devices can receive and use the updated model weights; through the introduction of policy tokens, the distribution and use of the model are finely controlled, enhancing security and reducing the risk of unauthorized access.

[0088] The cloud encapsulates the updated encrypted model weights, policy tokens, and metadata into a downlink data packet and transmits it to each end device via a self-healing multi-hop mesh link. Upon receiving the downlink data packet, the end device first verifies its own authority using the policy token. It then decrypts the updated encrypted model weights using the key agreed upon with the cloud, obtaining the plaintext model weights for the next round of object detection.

[0089] The use of self-healing multi-hop mesh links to transmit downlink data packets can ensure the reliability of data transmission in scenarios where the network is unstable or blocked.

[0090] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0091] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0093] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0094] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A portable wireless night vision monitoring system, characterized by: include, Utilize the trusted execution environment to instantly generate one-time session keys and encrypt low-light raw image frames in real time, so that the data is in a ciphertext state from the moment it is generated; On the terminal device, the fully homomorphic encryption framework and the quantized and slimmed-down YOLONano model are used to perform target detection on the ciphertext image frame, and the encrypted target box and confidence stream are output; The leakage sensitivity index and convergence contribution index are calculated in the terminal's trusted execution environment. The differential noise coefficient is then input into the privacy-utility trade-off model to adaptively adjust the Laplace noise scale. Local gradients are then noised and quantized to generate uplink data packets. The multi-band KineticMesh channel encapsulates and transmits data packets, monitors channel quality in real time, and automatically updates the routing table to achieve self-healing forwarding, ensuring that data packets reach the monitoring center. After the cloud verifies the signature of the data packet uploaded by the terminal device, it aggregates the noisy quantized gradients of multiple terminals to update the encryption model weights, and generates policy tokens and compliance metadata, which are then sent to the terminal device.

2. The portable wireless night vision monitoring system according to claim 1, characterized in that: The monitoring device detects the movement of the target object through the motion data acquisition device or starts the image acquisition process according to the preset time plan; when the image acquisition process is started, the terminal device uses a random number generator in the trusted execution environment to generate a 256-bit one-time session key.

3. The portable wireless night vision monitoring system according to claim 2, characterized in that: The monitoring device generates low-light original image frames through infrared fill light and low-light imaging technology, and uses a one-time session key and a randomly generated initialization vector to perform block encryption of the low-light original image frames with an Advanced Encryption Standard 256-bit key length to generate ciphertext image frames; After encryption is completed, the terminal device immediately destroys the one-time session key from the trusted execution environment and uses a hash-based message authentication code algorithm to generate a message authentication code for the ciphertext image frame to perform verification on the ciphertext image frame.

4. The portable wireless night vision monitoring system according to claim 3, characterized in that: The terminal device deploys a fully homomorphic encryption framework to convert the obtained encrypted low-light original image frame, i.e., the ciphertext image frame, into a ciphertext format compatible with the fully homomorphic encryption framework; The lightweight target detection model YOLONano is selected and 8-bit integer quantization is performed on it to generate quantized model parameters; the model's activation function is approximated by Chebyshev polynomials instead of the original nonlinear activation function to adapt to the operational limitations of the fully homomorphic encryption framework.

5. The mobile wireless night vision monitoring system according to claim 4, characterized in that: The quantized YOLONano model is used to perform homomorphic convolution feature extraction, homomorphic pooling, target localization and classification on the converted ciphertext image frame in the fully homomorphic encryption framework, and outputs the encrypted target box and confidence stream. The input ciphertext image frame is processed in blocks, combined with the parallel processing of model quantization and fully homomorphic encryption framework.

6. The portable wireless night vision monitoring system according to claim 5, characterized in that: In a trusted execution environment, the terminal device uses the encrypted target frame and confidence flow to calculate the local gradient, calculates the Fisher information matrix trace based on the local gradient, and performs virtual reconstruction to obtain the average peak signal-to-noise ratio of the reconstructed frame and the encrypted original frame after decryption, and calculates the leakage sensitivity index based on this.

7. The portable wireless night vision monitoring system according to claim 6, characterized in that: The terminal device obtains the historical global gradient sent by the cloud and calculates the cosine similarity between the local gradient and the historical global gradient; The instantaneous loss reduction rate after applying the local gradient update model is calculated on the validation set of the terminal device, and the convergence contribution index is calculated based on this.

8. The portable wireless night vision monitoring system according to claim 7, characterized in that: The terminal device inputs the leakage sensitivity index and convergence contribution index into the pre-trained privacy-utility balance model, which outputs the differential noise coefficient. The Laplace noise scale is calculated based on the differential noise coefficient, the global sensitivity of the local gradient, and the differential privacy budget. Laplace noise is added to each component of the local gradient and quantized to generate a noisy gradient. The terminal device encapsulates the quantized noisy gradient set into an uplink data packet and adds the terminal device signature.

9. The portable wireless night vision monitoring system according to claim 8, characterized in that: When transmitting data packets, the terminal device sends beacon frames at regular intervals to monitor signal strength, signal-to-noise ratio, and packet loss rate, and calculate channel quality indicators; When the channel quality index is lower than the preset threshold, the terminal device selects the path with the highest channel quality index or switches to the backup frequency band according to the routing table, updates the routing table and retransmits the data packet using the new path or new frequency band; After receiving the transmission data packet, the monitoring center sends a confirmation frame. If the terminal device does not receive the confirmation frame within the specified time, it resends the transmission data packet. When the upper limit of retransmission times is exceeded, an alarm is triggered.

10. The mobile wireless night vision monitoring system according to claim 9, characterized in that: The data processing unit obtains the real-time data collected by the data acquisition device, analyzes the real-time data based on a preset algorithm, generates analysis results, and transmits the analysis results to the display module for real-time display; Among them, the data acquisition device continuously monitors the target parameters, the data processing unit performs data processing at fixed time intervals, and the display module synchronously updates the output of the analysis results.

Images