An AI-based router, gateway, camera data protection method and system

Abstract

The application provides an AI-based router, gateway and camera data protection method and system in the technical field of Internet of Things and intelligent security, which comprises the following steps: S1, receiving a desensitization strategy configured by a user through a gateway, and storing the desensitization strategy in a local; the desensitization strategy comprises a private information type to be desensitized, a desensitization mode, a desensitization strength and an effective scene; S2, transmitting original video streams collected by a camera to a gateway through an encryption channel constructed by a router; S3, using a pre-deployed AI model to identify private information in the original video streams in real time, and performing desensitization processing on the identified private information according to the desensitization strategy to generate desensitized video streams; and S4, outputting the desensitized video streams for storage, display or sharing. The application has the advantages that the collaboration, real-time performance and compliance of data protection are greatly improved, and the network bandwidth consumption and data leakage risk are reduced.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things and smart security technology, and in particular to an AI-based method and system for protecting router, gateway, and camera data. Background Technology

[0002] With the rapid development of IoT technology and the widespread application of intelligent security monitoring systems, cameras have become core devices for home security, office management, and public area monitoring. Routers and gateways, as the data transmission hubs in IoT systems, are responsible for forwarding, storing, and scheduling video data. However, video streams captured by cameras typically contain a large amount of sensitive and private information, such as faces, license plates, ID numbers, and home addresses. If this data is not securely processed during transmission, storage, or sharing, it is highly susceptible to illegal theft, alteration, or misuse, leading to serious privacy breaches.

[0003] Currently, the mainstream privacy protection solutions on the market mainly include the following: 1. Centralized Cloud-Based Desensitization: The entire original video data is uploaded to a cloud server, where cloud computing power is used to identify and desensitize privacy information before returning the processed data. This method suffers from high data transmission latency and high bandwidth consumption, and there is a risk of leakage of the original data during transmission, which does not comply with the principle of transmitting only the minimum necessary data.

[0004] 2. Local fixed desensitization of cameras: Fixed desensitization rules are preset on the camera end (such as blurring faces in fixed areas), but it can only handle preset scenarios, has poor flexibility, and cannot adapt to dynamically changing privacy information (such as moving faces or license plates).

[0005] 3. Manual post-processing: Manually screening and de-identifying stored videos is extremely inefficient, cannot meet the needs of real-time monitoring, and is prone to human error.

[0006] 4. Simple Obscuring Tool: The software provides basic blurring or obscuring functions, but it requires manual operation, cannot intelligently identify dynamically appearing privacy information, and has insufficient adaptability.

[0007] The aforementioned existing technologies have obvious drawbacks: 1. Incomplete privacy protection: Traditional methods have difficulty accurately identifying various types of privacy information that appear dynamically in videos (such as temporarily appearing faces or moving license plates), resulting in limited coverage of desensitization.

[0008] 2. Insufficient real-time performance: The centralized cloud processing mode cannot meet the needs of real-time monitoring and linkage due to high data transmission and processing latency, and it also consumes a large amount of network bandwidth.

[0009] 3. High compliance risks: Raw video data may be intercepted or leaked during transmission to the cloud, which does not comply with the requirements of data localization and privacy protection regulations.

[0010] 4. Poor flexibility and adaptability: Fixed desensitization rules cannot adapt to different scenarios (such as home, office, public areas) or the desensitization needs of new privacy information, and lack configurability.

[0011] 5. Lack of coordinated management: Network devices such as routers and gateways are not involved in the unified management of de-identification strategies, resulting in inconsistent de-identification rules among multiple camera devices, making it difficult to achieve global privacy protection.

[0012] Therefore, how to provide an AI-based method and system for protecting router, gateway, and camera data, thereby improving the synergy, real-time performance, and compliance of data protection while reducing network bandwidth consumption and the risk of data leakage, has become an urgent technical problem to be solved. Summary of the Invention

[0013] The technical problem to be solved by this invention is to provide an AI-based method and system for protecting router, gateway, and camera data, thereby improving the coordination, real-time performance, and compliance of data protection, while reducing network bandwidth consumption and the risk of data leakage.

[0014] In a first aspect, the present invention provides an AI-based method for protecting router, gateway, and camera data, comprising the following steps: Step S1: Receive the user-configured de-identification policy through the gateway and store the de-identification policy locally; the de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario; Step S2: The camera transmits the raw video stream it captures to the gateway via an encrypted channel built by the router. Step S3: The gateway uses a pre-deployed AI model to identify privacy information in the original video stream in real time, and performs desensitization processing on the identified privacy information according to the desensitization strategy to generate a desensitized video stream; Step S4: The gateway outputs the de-identified video stream for storage, display, or sharing; Step S5: The gateway records a desensitization process log in real time, including at least the desensitization time, device number, privacy information type, desensitization method, and data flow direction, and periodically conducts spot checks on the desensitized video stream to verify the desensitization effect.

[0015] Furthermore, step S1 specifically includes: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario, and then send the desensitization strategy to the gateway. The gateway receives the de-identification policy and stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

[0016] Furthermore, in step S2, the encrypted channel adopts the TLS 1.3 protocol, and ensures that the original video stream without anonymization is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

[0017] Furthermore, in step S3, the AI ​​model is a lightweight multi-target detection and classification model, which is implemented based on a combination of YOLOv8-Tiny and MobileNetV3. It is used to identify at least one type of privacy information, including faces, license plates, ID card numbers, bank card numbers, sensitive text, and user-defined sensitive targets, and to dynamically track moving privacy targets to achieve continuous desensitization. The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

[0018] Furthermore, in step S3, the desensitization process employs an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames.

[0019] Secondly, the present invention provides an AI-based router, gateway, and camera data protection system, comprising the following modules: The de-identification policy receiving module is used to receive the de-identification policy configured by the user through the gateway and store the de-identification policy locally; the de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario; The raw video stream acquisition module is used to transmit the raw video stream acquired by the camera to the gateway via an encrypted channel built by the router. The desensitization processing module is used by the gateway to identify privacy information in the original video stream in real time using a pre-deployed AI model, and to desensitize the identified privacy information according to the desensitization strategy to generate a desensitized video stream; The desensitized video stream output module is used by the gateway to output the desensitized video stream for storage, display, or sharing; The de-identification process log recording module is used by the gateway to record in real time a de-identification process log that includes at least the de-identification time, device number, privacy information type, de-identification method, and data flow direction, and to periodically conduct spot checks on the de-identified video stream to verify the de-identification effect.

[0020] Furthermore, the desensitization strategy receiving module is specifically used for: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario, and then send the desensitization strategy to the gateway. The gateway receives the de-identification policy and stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

[0021] Furthermore, in the original video stream acquisition module, the encrypted channel adopts the TLS 1.3 protocol, and ensures that the un-de-identified original video stream is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

[0022] Furthermore, in the desensitization processing module, the AI ​​model is a lightweight multi-target detection and classification model, implemented based on a combination of YOLOv8-Tiny and MobileNetV3, used to identify at least one type of privacy information including faces, license plates, ID card numbers, bank card numbers, sensitive text, and user-defined sensitive targets, and to dynamically track moving privacy targets to achieve continuous desensitization; The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

[0023] Furthermore, in the desensitization processing module, the desensitization processing adopts an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames.

[0024] The advantages of this invention are: 1. The gateway receives and stores the user-configured desensitization policy, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario. The camera transmits the raw video stream it captures to the gateway via an encrypted channel built by the router. The gateway uses an AI model to identify privacy information in the raw video stream in real time. Based on the desensitization policy, it desensitizes the identified privacy information to generate a desensitized video stream, which is then output for storage, display, or sharing. The gateway records a desensitization process log in real time, including at least the desensitization time, device number, privacy information type, desensitization method, and data flow direction. It also periodically performs spot checks on the desensitized video stream to verify the desensitization effect. In other words, by building a collaborative architecture with the gateway as the intelligent control center, the gateway uniformly receives and manages the user-configured desensitization policy. The system coordinates multiple cameras within the local area network (LAN) via a router, ensuring consistency and global collaboration in privacy protection strategies. Regarding real-time performance, a lightweight AI model is deployed on the gateway, enabling real-time identification and de-identification of privacy information on the raw video stream locally. Efficiency is further enhanced through inter-frame optimization algorithms, completely avoiding the high latency and network bandwidth consumption associated with uploading the raw video stream to the cloud. In terms of compliance and security, the system requires that the raw video stream from the cameras be transmitted only through an encrypted channel built by the router within the LAN, strictly prohibiting uploading to the cloud. This achieves localized processing of sensitive data, fundamentally eliminating the risk of data leakage during public network transmission. Ultimately, this significantly improves the collaboration, real-time performance, and compliance of data protection, while reducing network bandwidth consumption and the risk of data leakage.

[0025] 2. By integrating an AI model through the gateway, privacy information (such as faces, license plates, etc.) in the original video stream is identified in real time, and processed immediately according to the user-configured de-identification strategy to generate a de-identified video stream. This real-time de-identification mechanism ensures that privacy data is effectively protected before storage, display, or sharing, significantly reducing the risk of privacy leakage. Compared with traditional post-processing or simple masking methods, this solution combines the intelligent recognition capabilities of AI to dynamically track moving targets and achieve continuous de-identification, thereby providing more comprehensive and proactive privacy protection in video surveillance scenarios.

[0026] 3. Users can customize the de-identification strategy through a visual interface, including the type of privacy information, de-identification method, strength, and effective scenarios. The gateway performs version control and dynamic updates of the de-identification strategy, allowing users to adjust the settings according to specific needs. This flexibility enables the solution to adapt to diverse application environments, such as home security, public place monitoring, or commercial use, while improving the user experience, avoiding the limitations of a "one-size-fits-all" approach, and enhancing the practicality and universality of the solution.

[0027] 4. By adopting lightweight AI models (such as a combination of YOLOv8-Tiny and MobileNetV3) and inter-frame optimization algorithms, redundant computation between adjacent frames is reduced, enabling real-time processing without the need for high-performance hardware support. This not only reduces the processing load on the gateway but also ensures stable operation of the system on resource-constrained devices. Compared with traditional heavy AI models, this design improves efficiency while maintaining high accuracy, making it suitable for large-scale deployment and offering significant technical advantages and economic benefits.

[0028] 5. The original video stream is transmitted through an encrypted channel built with the TLS 1.3 protocol and is prohibited from being uploaded to the cloud, ensuring the confidentiality and integrity of the data during transmission. This design strengthens local control and reduces the risk of external attacks from cloud storage. The gateway acts as a control center to uniformly manage multiple cameras, further improving the security and manageability of the system, which is superior to solutions that rely on cloud services.

[0029] 6. Real-time recording of the de-identification process log (including time, device number, privacy type, etc.) and regular random checks of the de-identified video streams to verify the effect. This provides a complete audit trail, making it easy to trace problems, optimize strategies, and meet compliance requirements. This built-in self-verification mechanism enhances the reliability of the solution, reduces the probability of incorrect or missed de-identification, thereby improving user trust and the product's market competitiveness.

[0030] 7. It innovatively integrates AI recognition, encrypted transmission, and policy management into a collaborative architecture of routers, gateways, and cameras, achieving end-to-end privacy protection. Compared to fragmented solutions, this integrated design simplifies the deployment process and reduces maintenance costs. At the same time, the AI ​​model supports various privacy information recognition and custom de-identification styles, demonstrating the foresight and adaptability of the technology.

[0031] 8. By integrating AI models to identify and process privacy information in video streams in real time, comprehensive and efficient privacy protection is achieved. Its advantages include highly customizable de-identification strategies to adapt to diverse user needs, lightweight AI and inter-frame optimization algorithms to ensure efficient resource utilization, encrypted channels and localized controls to enhance data security, auditable logs and verification mechanisms to improve reliability and compliance, and innovative technology integration to provide end-to-end solutions, thus demonstrating significant practicality, security and competitive advantages in the field of intelligent security. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Figure 1 This is a flowchart of an AI-based method for protecting router, gateway, and camera data according to the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of an AI-based router, gateway, and camera data protection system according to the present invention. Detailed Implementation

[0035] The overall approach of the technical solution in this application is as follows: By constructing a collaborative architecture with a gateway as the intelligent control center, the gateway uniformly receives and manages the desensitization policies configured by the user, and coordinates multiple cameras within the local area network through the router, ensuring the consistency and global collaboration of privacy protection policies; In terms of real-time performance, a lightweight AI model is deployed on the gateway, enabling real-time identification and desensitization of privacy information in the original video stream locally, and efficiency is improved through inter-frame optimization algorithms, completely avoiding the high latency and network bandwidth consumption caused by uploading the original video stream to the cloud; In terms of compliance and security, the original video stream from the camera is required to be transmitted only within the local area network through an encrypted channel built by the router, and uploading to the cloud is strictly prohibited, realizing the localized processing of sensitive data, fundamentally cutting off the risk of leakage of original data during transmission over the public network, thereby improving the collaboration, real-time performance, and compliance of data protection, while reducing network bandwidth consumption and the risk of data leakage.

[0036] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the AI-based router, gateway, and camera data protection method of the present invention includes the following steps: Step S1: Receive the user-configured de-identification policy through the gateway and store the de-identification policy locally; the de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario (e.g., public area monitoring requires de-identification of faces, while home internal monitoring only requires de-identification of stranger faces). Step S2: The camera (IP camera, Wi-Fi camera, PoE camera) transmits the raw video stream it captures to the gateway via an encrypted channel built by the router; the router establishes a network connection between the camera and the gateway through a local area network (Wi-Fi / Ethernet / Zigbee) to ensure the stability of data transmission. Step S3: The gateway uses a pre-deployed AI model to identify privacy information in the original video stream in real time. According to the desensitization strategy, the identified privacy information is desensitized to generate a desensitized video stream. In specific implementation, the desensitization intensity is adaptively adjustable, and the desensitization degree can be adjusted according to the scenario requirements (such as public display or internal storage). Step S4: The gateway outputs the de-identified video stream for storage, display, or sharing; The anonymized video stream generated by the gateway (re-encoded into a standard video format) no longer contains sensitive privacy information, therefore it can be securely used for various purposes: 1. Storage: It can be securely stored on local network attached storage or a home cloud drive, and even uploaded to the cloud for long-term backup or analysis under certain compliance requirements, without any privacy concerns. 2. Display: Live streaming or playback can be performed on mobile apps, computer clients, or monitoring screens. 3. Sharing: It can be securely provided to third parties, such as property management companies, law enforcement agencies (under legal conditions), or family members for remote viewing.

[0037] Step S5: The gateway records a desensitization process log in real time, including at least the desensitization time, device number, privacy information type, desensitization method, and data flow direction, and periodically conducts spot checks on the desensitized video stream to verify the desensitization effect.

[0038] To meet compliance requirements and ensure system reliability, the gateway records detailed de-identification process logs in real time. Log entries include at least: timestamp, camera device ID, type of privacy information identified, application de-identification method, coordinates of the de-identified area, and data flow (e.g., "stored to NAS," "pushed to App"). These logs are encrypted and stored, and can be queried and exported by authorized users for auditing and tracing purposes.

[0039] In addition, the system will automatically perform random checks on the generated de-identified video streams periodically (e.g., once a day) or according to rules (e.g., every 1000 frames processed). The random check process may involve re-identifying the sampled frames using an AI model to verify whether any privacy information has still been missed; or it may involve using an auxiliary, lighter-weight quality inspection model to assess whether the visual effect of the de-identified areas meets the preset intensity. The random check results will generate a report. If any anomalies are found (e.g., missed de-identification, poor de-identification effect), the system can issue an alarm and prompt the administrator to check or update the de-identification strategy.

[0040] Step S1 specifically involves: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario. The desensitization strategy is then sent to the gateway after being encrypted with AES-256. The gateway decrypts and receives the de-identification policy, stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

[0041] The de-identification strategy is a structured data set, including at least: Types of privacy information that need to be desensitized: These can be enumerated and selected, such as faces, license plates, ID cards, bank cards, and sensitive text (such as house numbers and document numbers). User-defined types are also supported (by uploading a small number of sample images for the system to fine-tune the model or perform feature matching).

[0042] Desensitization methods: Different algorithms can be configured for different types. For example, Gaussian blur, pixelation, or mosaic processing can be used for faces; rectangular masking, character replacement (such as replacing with *), or local distortion can be used for license plates.

[0043] Desensitization Intensity: This is a quantifiable parameter. For example, for Gaussian blur, the blur radius can be set (e.g., 5-15 pixels); for pixelation, the pixel block size can be set. It supports automatic adjustment of intensity based on the application scenario. For example, high-intensity desensitization is used in "public area monitoring" scenarios, low-intensity desensitization or full desensitization is used for family members' faces in "home monitoring" scenarios, while high-intensity desensitization is still used for strangers' faces.

[0044] Effective scenarios: Triggers can be based on conditions such as time, camera location, and image content (e.g., determined by scene recognition models to be a living room, doorway, or street). For example, it can be set to "de-sensitize faces captured by cameras in the office area only on weekdays from 9:00 to 18:00".

[0045] Users customize the above policies through visual interfaces such as web or mobile apps. After receiving the policies, the gateway not only stores them in local non-volatile memory, but also performs version control (recording the historical versions of each modification for easy rollback) and dynamic updates (new policies can be issued and take effect immediately or on schedule without restarting the device). As the control center, the gateway broadcasts unified policy instructions to all cameras in the local area network through the router, ensuring policy consistency.

[0046] In step S2, the encrypted channel adopts the TLS 1.3 protocol and ensures that the original video stream without anonymization is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

[0047] Cameras (including but not limited to IP cameras, Wi-Fi cameras, and PoE cameras) capture raw video streams (e.g., in H.264 / H.265 format). The router, acting as the network hub, is responsible for establishing an end-to-end encrypted channel between the cameras and the gateway. Preferably, this encrypted channel uses TLS 1.3 or a higher version protocol, leveraging its forward security and other advantages to ensure that the video stream is not eavesdropped on or tampered with during transmission over the local area network. Crucially, the system design ensures that raw video stream data containing sensitive information is transmitted only within this encrypted channel and is never uploaded to any cloud server. This strictly adheres to the principle of data minimization, reducing the risk of data leakage at its source.

[0048] In step S3, the AI ​​model is a lightweight multi-object detection and classification model, implemented based on a combination of YOLOv8-Tiny and MobileNetV3. It is used to identify at least one type of privacy information, including faces, license plates, ID card numbers, bank card numbers, sensitive text (such as house numbers), and user-defined sensitive targets. It also dynamically tracks moving privacy targets to achieve continuous desensitization. In specific implementations, it also supports the customization of privacy information types, allowing users to add sensitive targets for specific scenarios (such as confidential company documents or home safes). The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

[0049] In step S3, the desensitization process employs an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames and improve processing efficiency.

[0050] The gateway, as the processing core, pre-deploys a lightweight multi-object detection and classification AI model. This model is preferably based on a combination architecture of YOLOv8-Tiny (responsible for fast object localization) and MobileNetV3 (responsible for efficient feature extraction and classification), and undergoes targeted training and optimization to achieve real-time processing within the gateway's limited computing power (such as ARM Cortex-A series processors). This model can identify various privacy-sensitive information and dynamically track moving targets (such as walking people and moving vehicles), ensuring continuous and seamless desensitization throughout their appearance in the video frame.

[0051] In practice, the gateway decodes the received raw video stream and feeds it into the AI ​​model for recognition frame by frame or as needed. After identifying the privacy region (usually represented by bounding box coordinates), the gateway calls the corresponding de-identification algorithm library (such as image processing functions in OpenCV) to process the specified region according to the de-identification strategy.

[0052] To improve processing efficiency, the desensitization process employs an inter-frame optimization algorithm. Specifically, it can utilize the temporal continuity of the video: For targets moving slowly in a static background, the detection frequency can be reduced (e.g., full detection every 5 frames). In intermediate frames, target tracking algorithms (e.g., KCF, SORT) can be used to predict and update the desensitized area based on the position of the previous frame.

[0053] For privacy information whose position and shape do not change much in consecutive frames (such as license plates fixed on a wall), the desensitized area of ​​the previous frame can be reused, and re-identification is only performed when the scene changes significantly (determined by optical flow or frame difference method).

[0054] This optimization significantly reduces computational load, ensuring the system maintains low latency even with high-resolution, high-frame-rate video streams.

[0055] A preferred embodiment of the AI-based router, gateway, and camera data protection system of the present invention includes the following modules: The de-identification policy receiving module is used to receive the de-identification policy configured by the user through the gateway and store the de-identification policy locally. The de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario (such as de-identifying faces for public area monitoring and de-identifying stranger faces only for home internal monitoring). The raw video stream acquisition module is used by cameras (IP cameras, Wi-Fi cameras, PoE cameras) to transmit the raw video streams they acquire to the gateway via an encrypted channel built by the router; the router establishes a network connection between the camera and the gateway through a local area network (Wi-Fi / Ethernet / Zigbee) to ensure the stability of data transmission; The desensitization module is used by the gateway to identify privacy information in the original video stream in real time using a pre-deployed AI model, and to desensitize the identified privacy information according to the desensitization strategy to generate a desensitized video stream. In specific implementation, it supports the ability to adaptively adjust the desensitization intensity, and can adjust the desensitization degree according to the needs of the scenario (such as public display or internal storage). The desensitized video stream output module is used by the gateway to output the desensitized video stream for storage, display, or sharing; The anonymized video stream generated by the gateway (re-encoded into a standard video format) no longer contains sensitive privacy information, therefore it can be securely used for various purposes: 1. Storage: It can be securely stored on local network attached storage or a home cloud drive, and even uploaded to the cloud for long-term backup or analysis under certain compliance requirements, without any privacy concerns. 2. Display: Live streaming or playback can be performed on mobile apps, computer clients, or monitoring screens. 3. Sharing: It can be securely provided to third parties, such as property management companies, law enforcement agencies (under legal conditions), or family members for remote viewing.

[0056] The de-identification process log recording module is used by the gateway to record in real time a de-identification process log that includes at least the de-identification time, device number, privacy information type, de-identification method, and data flow direction, and to periodically conduct spot checks on the de-identified video stream to verify the de-identification effect.

[0057] To meet compliance requirements and ensure system reliability, the gateway records detailed de-identification process logs in real time. Log entries include at least: timestamp, camera device ID, type of privacy information identified, application de-identification method, coordinates of the de-identified area, and data flow (e.g., "stored to NAS," "pushed to App"). These logs are encrypted and stored, and can be queried and exported by authorized users for auditing and tracing purposes.

[0058] In addition, the system will automatically perform random checks on the generated de-identified video streams periodically (e.g., once a day) or according to rules (e.g., every 1000 frames processed). The random check process may involve re-identifying the sampled frames using an AI model to verify whether any privacy information has still been missed; or it may involve using an auxiliary, lighter-weight quality inspection model to assess whether the visual effect of the de-identified areas meets the preset intensity. The random check results will generate a report. If any anomalies are found (e.g., missed de-identification, poor de-identification effect), the system can issue an alarm and prompt the administrator to check or update the de-identification strategy.

[0059] The desensitization strategy receiving module is specifically used for: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario. The desensitization strategy is then sent to the gateway after being encrypted with AES-256. The gateway decrypts and receives the de-identification policy, stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

[0060] The de-identification strategy is a structured data set, including at least: Types of privacy information that need to be desensitized: These can be enumerated and selected, such as faces, license plates, ID cards, bank cards, and sensitive text (such as house numbers and document numbers). User-defined types are also supported (by uploading a small number of sample images for the system to fine-tune the model or perform feature matching).

[0061] Desensitization methods: Different algorithms can be configured for different types. For example, Gaussian blur, pixelation, or mosaic processing can be used for faces; rectangular masking, character replacement (such as replacing with *), or local distortion can be used for license plates.

[0062] Desensitization Intensity: This is a quantifiable parameter. For example, for Gaussian blur, the blur radius can be set (e.g., 5-15 pixels); for pixelation, the pixel block size can be set. It supports automatic adjustment of intensity based on the application scenario. For example, high-intensity desensitization is used in "public area monitoring" scenarios, low-intensity desensitization or full desensitization is used for family members' faces in "home monitoring" scenarios, while high-intensity desensitization is still used for strangers' faces.

[0063] Effective scenarios: Triggers can be based on conditions such as time, camera location, and image content (e.g., determined by scene recognition models to be a living room, doorway, or street). For example, it can be set to "de-sensitize faces captured by cameras in the office area only on weekdays from 9:00 to 18:00".

[0064] Users customize the above policies through visual interfaces such as web or mobile apps. After receiving the policies, the gateway not only stores them in local non-volatile memory, but also performs version control (recording the historical versions of each modification for easy rollback) and dynamic updates (new policies can be issued and take effect immediately or on schedule without restarting the device). As the control center, the gateway broadcasts unified policy instructions to all cameras in the local area network through the router, ensuring policy consistency.

[0065] In the original video stream acquisition module, the encrypted channel adopts the TLS 1.3 protocol, and ensures that the original video stream without anonymization is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

[0066] Cameras (including but not limited to IP cameras, Wi-Fi cameras, and PoE cameras) capture raw video streams (e.g., in H.264 / H.265 format). The router, acting as the network hub, is responsible for establishing an end-to-end encrypted channel between the cameras and the gateway. Preferably, this encrypted channel uses TLS 1.3 or a higher version protocol, leveraging its forward security and other advantages to ensure that the video stream is not eavesdropped on or tampered with during transmission over the local area network. Crucially, the system design ensures that raw video stream data containing sensitive information is transmitted only within this encrypted channel and is never uploaded to any cloud server. This strictly adheres to the principle of data minimization, reducing the risk of data leakage at its source.

[0067] In the desensitization processing module, the AI ​​model is a lightweight multi-target detection and classification model, implemented based on a combination of YOLOv8-Tiny and MobileNetV3. It is used to identify at least one type of privacy information, including faces, license plates, ID card numbers, bank card numbers, sensitive text (such as house numbers), and user-defined sensitive targets. It also dynamically tracks moving privacy targets to achieve continuous desensitization. In specific implementations, it also supports the customization of privacy information types, allowing users to add sensitive targets for specific scenarios (such as confidential company documents or home safes). The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

[0068] In the desensitization processing module, the desensitization processing adopts an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames and improve processing efficiency.

[0069] The gateway, as the processing core, pre-deploys a lightweight multi-object detection and classification AI model. This model is preferably based on a combination architecture of YOLOv8-Tiny (responsible for fast object localization) and MobileNetV3 (responsible for efficient feature extraction and classification), and undergoes targeted training and optimization to achieve real-time processing within the gateway's limited computing power (such as ARM Cortex-A series processors). This model can identify various privacy-sensitive information and dynamically track moving targets (such as walking people and moving vehicles), ensuring continuous and seamless desensitization throughout their appearance in the video frame.

[0070] In practice, the gateway decodes the received raw video stream and feeds it into the AI ​​model for recognition frame by frame or as needed. After identifying the privacy region (usually represented by bounding box coordinates), the gateway calls the corresponding de-identification algorithm library (such as image processing functions in OpenCV) to process the specified region according to the de-identification strategy.

[0071] To improve processing efficiency, the desensitization process employs an inter-frame optimization algorithm. Specifically, it can utilize the temporal continuity of the video: For targets moving slowly in a static background, the detection frequency can be reduced (e.g., full detection every 5 frames). In intermediate frames, target tracking algorithms (e.g., KCF, SORT) can be used to predict and update the desensitized area based on the position of the previous frame.

[0072] For privacy information whose position and shape do not change much in consecutive frames (such as license plates fixed on a wall), the desensitized area of ​​the previous frame can be reused, and re-identification is only performed when the scene changes significantly (determined by optical flow or frame difference method).

[0073] This optimization significantly reduces computational load, ensuring the system maintains low latency even with high-resolution, high-frame-rate video streams.

[0074] In summary, the advantages of this invention are as follows: 1. The gateway receives and stores the user-configured desensitization policy, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario. The camera transmits the raw video stream it captures to the gateway via an encrypted channel built by the router. The gateway uses an AI model to identify privacy information in the raw video stream in real time. Based on the desensitization policy, it desensitizes the identified privacy information to generate a desensitized video stream, which is then output for storage, display, or sharing. The gateway records a desensitization process log in real time, including at least the desensitization time, device number, privacy information type, desensitization method, and data flow direction. It also periodically performs spot checks on the desensitized video stream to verify the desensitization effect. In other words, by building a collaborative architecture with the gateway as the intelligent control center, the gateway uniformly receives and manages the user-configured desensitization policy. The system coordinates multiple cameras within the local area network (LAN) via a router, ensuring consistency and global collaboration in privacy protection strategies. Regarding real-time performance, a lightweight AI model is deployed on the gateway, enabling real-time identification and de-identification of privacy information on the raw video stream locally. Efficiency is further enhanced through inter-frame optimization algorithms, completely avoiding the high latency and network bandwidth consumption associated with uploading the raw video stream to the cloud. In terms of compliance and security, the system requires that the raw video stream from the cameras be transmitted only through an encrypted channel built by the router within the LAN, strictly prohibiting uploading to the cloud. This achieves localized processing of sensitive data, fundamentally eliminating the risk of data leakage during public network transmission. Ultimately, this significantly improves the collaboration, real-time performance, and compliance of data protection, while reducing network bandwidth consumption and the risk of data leakage.

[0075] 2. By integrating an AI model through the gateway, privacy information (such as faces, license plates, etc.) in the original video stream is identified in real time, and processed immediately according to the user-configured de-identification strategy to generate a de-identified video stream. This real-time de-identification mechanism ensures that privacy data is effectively protected before storage, display, or sharing, significantly reducing the risk of privacy leakage. Compared with traditional post-processing or simple masking methods, this solution combines the intelligent recognition capabilities of AI to dynamically track moving targets and achieve continuous de-identification, thereby providing more comprehensive and proactive privacy protection in video surveillance scenarios.

[0076] 3. Users can customize the de-identification strategy through a visual interface, including the type of privacy information, de-identification method, strength, and effective scenarios. The gateway performs version control and dynamic updates of the de-identification strategy, allowing users to adjust the settings according to specific needs. This flexibility enables the solution to adapt to diverse application environments, such as home security, public place monitoring, or commercial use, while improving the user experience, avoiding the limitations of a "one-size-fits-all" approach, and enhancing the practicality and universality of the solution.

[0077] 4. By adopting lightweight AI models (such as a combination of YOLOv8-Tiny and MobileNetV3) and inter-frame optimization algorithms, redundant computation between adjacent frames is reduced, enabling real-time processing without the need for high-performance hardware support. This not only reduces the processing load on the gateway but also ensures stable operation of the system on resource-constrained devices. Compared with traditional heavy AI models, this design improves efficiency while maintaining high accuracy, making it suitable for large-scale deployment and offering significant technical advantages and economic benefits.

[0078] 5. The original video stream is transmitted through an encrypted channel built with the TLS 1.3 protocol and is prohibited from being uploaded to the cloud, ensuring the confidentiality and integrity of the data during transmission. This design strengthens local control and reduces the risk of external attacks from cloud storage. The gateway acts as a control center to uniformly manage multiple cameras, further improving the security and manageability of the system, which is superior to solutions that rely on cloud services.

[0079] 6. Real-time recording of the de-identification process log (including time, device number, privacy type, etc.) and regular random checks of the de-identified video streams to verify the effect. This provides a complete audit trail, making it easy to trace problems, optimize strategies, and meet compliance requirements. This built-in self-verification mechanism enhances the reliability of the solution, reduces the probability of incorrect or missed de-identification, thereby improving user trust and the product's market competitiveness.

[0080] 7. It innovatively integrates AI recognition, encrypted transmission, and policy management into a collaborative architecture of routers, gateways, and cameras, achieving end-to-end privacy protection. Compared to fragmented solutions, this integrated design simplifies the deployment process and reduces maintenance costs. At the same time, the AI ​​model supports various privacy information recognition and custom de-identification styles, demonstrating the foresight and adaptability of the technology.

[0081] 8. By integrating AI models to identify and process privacy information in video streams in real time, comprehensive and efficient privacy protection is achieved. Its advantages include highly customizable de-identification strategies to adapt to diverse user needs, lightweight AI and inter-frame optimization algorithms to ensure efficient resource utilization, encrypted channels and localized controls to enhance data security, auditable logs and verification mechanisms to improve reliability and compliance, and innovative technology integration to provide end-to-end solutions, thus demonstrating significant practicality, security and competitive advantages in the field of intelligent security.

[0082] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for protecting router, gateway, and camera data based on AI, characterized in that: Includes the following steps: Step S1: Receive the user-configured de-identification policy through the gateway and store the de-identification policy locally; the de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario; Step S2: The camera transmits the raw video stream it captures to the gateway via an encrypted channel built by the router. Step S3: The gateway uses a pre-deployed AI model to identify privacy information in the original video stream in real time, and performs desensitization processing on the identified privacy information according to the desensitization strategy to generate a desensitized video stream; Step S4: The gateway outputs the de-identified video stream for storage, display, or sharing; Step S5: The gateway records a desensitization process log in real time, including at least the desensitization time, device number, privacy information type, desensitization method, and data flow direction, and periodically conducts spot checks on the desensitized video stream to verify the desensitization effect.

2. The AI-based method for protecting router, gateway, and camera data as described in claim 1, characterized in that: Step S1 specifically involves: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario, and then send the desensitization strategy to the gateway. The gateway receives the de-identification policy and stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

3. The AI-based method for protecting router, gateway, and camera data as described in claim 1, characterized in that: In step S2, the encrypted channel adopts the TLS 1.3 protocol and ensures that the original video stream without anonymization is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

4. The AI-based method for protecting router, gateway, and camera data as described in claim 1, characterized in that: In step S3, the AI ​​model is a lightweight multi-target detection and classification model, which is based on a combination of YOLOv8-Tiny and MobileNetV3. It is used to identify at least one type of privacy information, including faces, license plates, ID card numbers, bank card numbers, sensitive text, and user-defined sensitive targets, and to dynamically track moving privacy targets to achieve continuous desensitization. The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

5. The AI-based method for protecting router, gateway, and camera data as described in claim 1, characterized in that: In step S3, the desensitization process employs an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames.

6. An AI-based data protection system for routers, gateways, and cameras, characterized in that: Includes the following modules: The de-identification policy receiving module is used to receive the de-identification policy configured by the user through the gateway and store the de-identification policy locally; the de-identification policy includes the type of privacy information to be de-identified, the de-identification method, the de-identification strength, and the effective scenario; The raw video stream acquisition module is used to transmit the raw video stream acquired by the camera to the gateway via an encrypted channel built by the router. The desensitization processing module is used by the gateway to identify privacy information in the original video stream in real time using a pre-deployed AI model, and to desensitize the identified privacy information according to the desensitization strategy to generate a desensitized video stream; The desensitized video stream output module is used by the gateway to output the desensitized video stream for storage, display, or sharing; The de-identification process log recording module is used by the gateway to record in real time a de-identification process log that includes at least the de-identification time, device number, privacy information type, de-identification method, and data flow direction, and to periodically conduct spot checks on the de-identified video stream to verify the de-identification effect.

7. The AI-based router, gateway, and camera data protection system as described in claim 6, characterized in that: The desensitization strategy receiving module is specifically used for: Users can customize the desensitization strategy through a visual interface, including the type of privacy information to be desensitized, the desensitization method, the desensitization strength, and the effective scenario, and then send the desensitization strategy to the gateway. The gateway receives the de-identification policy and stores it locally, and performs version control and dynamic updates on the de-identification policy; The gateway acts as the control center, managing multiple cameras within the local area network through the router, ensuring that all raw video streams captured by the cameras undergo privacy protection processing based on the same desensitization strategy.

8. The AI-based router, gateway, and camera data protection system as described in claim 6, characterized in that: In the original video stream acquisition module, the encrypted channel adopts the TLS 1.3 protocol, and ensures that the original video stream without anonymization is transmitted only between the camera and the gateway, and is prohibited from being uploaded to the cloud.

9. The AI-based router, gateway, and camera data protection system as described in claim 6, characterized in that: In the desensitization processing module, the AI ​​model is a lightweight multi-target detection and classification model, which is based on a combination of YOLOv8-Tiny and MobileNetV3. It is used to identify at least one type of privacy information, including faces, license plates, ID card numbers, bank card numbers, sensitive text, and user-defined sensitive targets, and to dynamically track moving privacy targets to achieve continuous desensitization. The process of desensitizing the identified privacy information specifically includes at least one of the following methods: Apply Gaussian blur or pixelation to the face; The license plate was either masked with black or the characters were replaced. Sensitive text is handled using both blurring and obscuring. User-defined sensitive targets are anonymized using user-specified styles.

10. The AI-based router, gateway, and camera data protection system as described in claim 6, characterized in that: In the desensitization processing module, the desensitization processing adopts an inter-frame optimization algorithm to reduce redundant calculations between adjacent frames.

Images