A privacy data hierarchical cloaking and content filtering method in railway communication

Abstract

This invention relates to the intersection of railway information security and multimedia processing, specifically a method for graded masking and content filtering of privacy data in railway communications. By configuring multi-level user permission rules, it applies varying degrees of privacy processing to keyframe images output by sensitive target detection targets. This allows for accurate identification of sensitive targets and masking only them to preserve background information. The masking intensity is dynamically adjusted based on user permissions, achieving effective voice desensitization and accurate keyword detection. Simultaneously, it establishes a comprehensive operation log and reverse audit closed loop. The advantages of this invention are: achieving a balance between information preservation and accurate masking; a role-based dynamic masking mechanism that meets the information detail requirements of different users, significantly improving flexibility compared to static masking techniques; high comprehensiveness of voice processing; a reliable audit closed loop ensuring traceability of operations and controllability of decryption; and wide applicability, suitable for railway communication scenarios in various fields such as government affairs and emergency response.

Description

Technical Field

[0001] This invention relates to the intersection of railway information security and multimedia processing, and in particular to a method for graded masking and content filtering of privacy data in railway communications, applicable to the protection and processing of sensitive information in video and voice in cross-departmental information sharing scenarios such as railway management and government affairs. Background Technology

[0002] In railway communication scenarios, privacy protection technology is crucial for ensuring secure information sharing. Currently, existing technologies mainly include static masking, general encryption, and simple keyword filtering, but these technologies have significant limitations.

[0003] In video masking, traditional mosaic technology indiscriminately blurs sensitive targets such as faces and personal information. This not only obscures sensitive information but also leads to a significant loss of background environmental information (such as terrain features and personnel postures), severely impacting the efficiency of cross-departmental collaboration. Furthermore, most existing dynamic masking technologies rely on a single model for detection, often resulting in insufficient accuracy in identifying sensitive targets in complex scenarios such as backlighting or occluded objects, failing to meet practical application needs.

[0004] Regarding permission adaptation, there is a lack of an effective hierarchical processing mechanism, making it impossible to dynamically adjust the masking intensity based on different user permissions (such as ordinary personnel and commanders). This easily leads to situations of "over-masking" or "under-masking," which does not comply with the "least privilege" principle in the "Regulations on Security Management of Public Security Information Sharing," and poses security risks to information sharing.

[0005] In speech processing, existing speech anonymization technologies mostly employ full-process voice changing. While this method conceals the speaker's identity to some extent, it disrupts the coherence of the speech's semantics and affects the accurate transmission of information. Keyword filtering technology, on the other hand, relies on simple string matching, which has low coverage for detecting variations such as homophones and pinyin abbreviations, and cannot effectively filter sensitive information.

[0006] In terms of auditing mechanisms, most technologies lack operational traceability or reverse auditing capabilities, making it impossible to trace the concealment process. Once an information security issue occurs, it is difficult to identify the cause and responsibility, posing a significant compliance risk.

[0007] In summary, existing railway communication privacy protection technologies suffer from drawbacks such as low accuracy in identifying sensitive targets, significant loss of background information, poor permission adaptability, inadequate voice processing performance, and a lack of effective auditing mechanisms. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the prior art by providing a method for graded masking and content filtering of privacy data in railway communications. This method can accurately identify sensitive targets and mask them only to retain background information. It can dynamically adjust the masking intensity according to user permissions, achieve effective voice desensitization and accurate keyword detection, and establish a complete operation log and reverse audit closed loop. This improves cross-departmental collaboration efficiency and meets relevant regulatory requirements while ensuring information security.

[0009] The objective of this invention is achieved through the following technical solutions: A method for graded masking and content filtering of privacy data in railway communications, characterized by the following steps: A main processing server is deployed at the core node of the railway communication network. A sensitive target detection model is established, verified and activated on the main processing server. The sensitive target detection model is preset with multiple types of sensitive target features. Multi-level user permission rules are configured in the background of the main processing server, and each level of user permission rules has different privacy processing principles; The main processing server receives and processes video streams and / or audio streams, wherein: When the main processing server receives the video stream, it preprocesses the video stream and extracts keyframe images. The sensitive target detection model analyzes the preprocessed keyframe image, identifies the sensitive target features contained in the keyframe image, and tracks the sensitive targets containing the sensitive target features. The main processing server obtains the user permission level of the terminal user requesting to view the video stream according to the identity authentication interface, and calls the corresponding privacy processing principle to perform privacy processing on the key frame images. The processed key frame images and non-key frame images are then recombined to form the processed video stream. When the main processing server receives the voice stream, it preprocesses the voice stream and extracts the valid voice segments. The system extracts voiceprint features from valid speech segments, converts the voiceprint features into timbre, transcribes the timbre-converted speech stream into text in real time, compares the transcribed text with a preset dynamic sensitive keyword library, and if a sensitive keyword is detected that is contained in the dynamic sensitive keyword library, the corresponding speech segment in the speech stream is replaced with a prompt sound. The processed video and / or audio streams are sent to the end users who request to view them and recorded. The original received video and audio streams are encrypted before being stored. When a security audit is required, it is performed jointly by multiple authorized personnel with specific permissions. The main processing server verifies the identity and permissions of the authorized personnel and decrypts the target file through a security mechanism for the authorized personnel to view online in a controlled environment.

[0010] The multi-level user permission rules are divided into three levels of permissions, including: Level 1 permission: Apply Gaussian blur with a relatively large blur radius to the face area, and cover the specific ID number area with an opaque rectangle; Level 2 permissions: Apply Gaussian blur with a relatively small blur radius or edge-preserving blur to the face area, and use a semi-transparent overlay to cover specific ID number areas; Level 3 access control: Overlay a specific watermark on the face area, and fully display and add border annotations to a specific ID number area.

[0011] When the main processing server receives the video stream, it performs keyframe image extraction processing on the video stream at certain intervals and applies image enhancement algorithms to the extracted keyframe images.

[0012] For sensitive targets, mark their locations on keyframe images.

[0013] For moving sensitive targets, a target tracking algorithm is enabled to predict their position in subsequent keyframe images, and re-detection is performed when the prediction deviation exceeds a threshold, ensuring that moving sensitive targets are continuously locked in the video stream. Use an applied image fusion algorithm to process the boundary between the occluded area and the original background.

[0014] When the main processing server receives the voice stream, it applies a noise reduction algorithm to remove background noise from the voice stream and retains the main human voice frequency bands; when a silent segment is detected, it is marked as invalid voice and no further processing is performed.

[0015] When multiple authorized personnel operate together, physical private keys are used for triggering, and decryption is achieved by synthesizing multiple physical private keys in real time.

[0016] The filenames of the original received video and audio streams are obfuscated, and the generated log files are encrypted and stored. These log files are periodically or in real-time synchronized to nodes or systems with auditing privileges.

[0017] The log files are stored on the railway department's main processing server or a dedicated storage device.

[0018] The advantages of this invention are: 1) Achieving a balance between information preservation and precise masking: By using image segmentation and hierarchical blurring technology, while maintaining a high accuracy rate in masking sensitive targets, background information is preserved, effectively solving the information loss problem caused by traditional mosaic and significantly improving the efficiency of cross-departmental collaboration.

[0019] 2) The role-based dynamic masking mechanism can meet the needs of different users for information detail, and complies with the "on-demand sharing" principle in the "Regulations on Security Management of Public Security Information Sharing". It is much more flexible than static masking technology.

[0020] 3) The speech processing is highly comprehensive. Voiceprint desensitization hides the speaker's identity while preserving the intelligibility of the speech semantics. Keyword filtering detection achieves high coverage, significantly improving security and compliance compared to traditional speech filtering technologies.

[0021] 4) The audit closed loop is reliable, ensuring the traceability of operations and the controllability of decryption.

[0022] 5) In terms of scenario scalability, the model supports adding new sensitive target types (special identifiers, classified equipment) through incremental training, which is applicable to railway communication scenarios in multiple fields such as government affairs and emergency response. Detailed Implementation

[0023] The following examples further illustrate the features and other related characteristics of the present invention to facilitate understanding by those skilled in the art: Example: The method for graded blocking and content filtering of privacy data in railway communications in this example includes the following steps: (1) System initialization and configuration: First, a main processing server is deployed at the core node of the railway communication network, and an operating system (Linux / CentOS) is installed. Dedicated client software is then installed on the terminals of departments that need to share information (e.g., dispatch centers, station management departments). After the client starts, it automatically registers its device MAC address with the main processing server and binds authorized user identity information.

[0024] Next, the pre-trained sensitive target detection model (containing multiple target features such as faces, specific ID numbers, and specific identifiers) is uploaded to the main processing server for verification and activation.

[0025] Multi-level user permission rules can be configured in the main processing server backend: Level 1 (lowest level): Apply a strong Gaussian blur to the face area (blur radius of about 6-8 pixels), and use an opaque rectangle to cover specific ID number areas.

[0026] Level 2 (Medium) Permissions: Apply a weak Gaussian blur or edge-preserving blur to the face area (blur radius approximately 3-5 pixels), and use a semi-transparent overlay layer for specific ID number areas.

[0027] Level 3 access (highest access): Overlay a specific watermark on the face area ("Authorized View"), and fully display a specific ID number area but add a prominent border label.

[0028] Once the configured permission rules table is reviewed and locked, subsequent modifications require multiple authorizations.

[0029] In this embodiment, three levels of permissions—first, second, and third—are used for explanation. However, in actual use, more levels of user permission rules can be designed according to actual needs. For example, more permissions can be set up with the size of the fuzzy radius as a preset limit, or different levels of user permission rules can be set up with each actual privacy processing requirement as a boundary.

[0030] (2) Video stream reception and preprocessing: When an end user initiates a video sharing request, the main processing server receives the real-time video stream from the sender, and the server receives video data at a set frame rate. To improve processing efficiency, keyframe images are extracted and processed at certain intervals (e.g., every 1-2 frames), while non-keyframe images are temporarily stored.

[0031] Image enhancement algorithms (contrast enhancement) are applied to the extracted keyframe images, and lighting conditions (backlighting, low light) are optimized to ensure the accuracy of subsequent target detection.

[0032] (3) Video Sensitive Target Detection and Tracking: The preprocessed keyframe images are analyzed using an object detection model (based on the YOLO deep learning series). Predefined sensitive target categories (faces, specific ID numbers, specific identifiers, etc.) are identified within the keyframe images. For each detected sensitive target, its location is precisely marked on the image (using rectangular bounding box coordinates, but other shapes can also be used).

[0033] For moving sensitive targets (such as people walking), enable the target tracking algorithm (which can use Kalman filtering) to predict its position in subsequent frames, and re-detect when the prediction deviation exceeds a threshold to ensure that the target is continuously locked in the video stream.

[0034] (4) Real-time masking based on permissions: The main processing server obtains the permission level of the terminal user who is currently requesting to view the video stream through the authentication interface.

[0035] Based on the user's permission level, invoke the corresponding rule configured to match their permission level: For users with Level 1 privileges: apply a strong Gaussian blur to all detected faces, and apply an opaque rectangular overlay to specific ID numbers.

[0036] For users with level 2 access: apply weak Gaussian blur or edge-preserving blur to detected faces, and apply a semi-transparent overlay to specific ID numbers.

[0037] For users with Level 3 access: watermarks are overlaid on detected faces, and specific ID numbers are displayed in full after being marked with borders.

[0038] In this embodiment, an image fusion algorithm is applied to process the boundary between the occluded area and the original background, so that the occlusion effect transitions naturally and avoids obvious edge marks.

[0039] The processed keyframes and non-keyframes are then recombined to form the processed video stream. The video stream after face masking should ensure that unauthorized users cannot identify specific individuals.

[0040] (5) Voice stream reception and preprocessing: The main processing server synchronously receives the audio stream accompanying the video.

[0041] The noise reduction algorithm (spectral subtraction) is applied to remove background noise (wind noise, equipment noise) while retaining the main human voice frequency band (300-3400Hz). Silent segments (volume below a set threshold) are detected and marked as invalid speech, and no further processing is performed.

[0042] (6) Voice content desensitization and filtering: Voiceprint features are extracted from valid speech segments (using the MFCC algorithm).

[0043] The extracted voiceprint features are matched against a pre-defined "neutral voiceprint library," and the closest neutral voiceprint model is selected. This neutral voiceprint model is then used to perform voice transformation (voice alteration) on the original speech, replacing the original speaker's timbre while maintaining the original speech rate and intonation to ensure semantic intelligibility and clarity. The transformed speech stream is transcribed into text in real time (using the CTC algorithm). The transcribed text is then compared with a pre-defined dynamic sensitive keyword library (containing keywords and their common variations, such as homophones and pinyin abbreviations, to improve the coverage of sensitive keyword detection) (using the AC automata algorithm).

[0044] If a sensitive keyword is detected, the corresponding audio segment is immediately replaced with a prompt sound ("beep") in the audio stream, and the duration of the prompt sound matches the duration of the keyword pronunciation.

[0045] (7) Distribution and recording of processing results: The processed video and audio streams are sent to the end user who requests to view them.

[0046] If step 6 detects that a sensitive keyword has been triggered, an alarm message is sent to the system administrator terminal, including the trigger time, keyword content, and source device.

[0047] The main processing server generates detailed log records for each processing task, including: processing start / end timestamps, processed video / audio file identifiers, terminal device MAC addresses, the number of various sensitive targets detected, the application's masking rule level, the number of sensitive keywords triggered, operator identity identifiers, and permission verification results.

[0048] (8) Secure storage and auditing: The original received video and audio data are encrypted (using the AES algorithm) and then stored on the railway department's main processing server or dedicated storage device, with the filenames obfuscated.

[0049] The generated log files are stored after being encrypted (SHA-256 hash). The log files are periodically or in real-time synchronized to nodes or systems with auditing privileges (such as blockchain-based consortium blockchain nodes).

[0050] When a security audit is required, it must be conducted jointly by multiple authorized personnel with specific high-level privileges (Level 3). Specifically, the audit action is triggered by personnel from two specific railway departments (Safety Supervision Section + Information Section) using a physical USB-Key.

[0051] Authorized personnel submit audit requests that include the reasons for the audit and document identification.

[0052] The system verifies the identity and permissions of authorized personnel. The target file is decrypted using a security mechanism (Shamir secret sharing algorithm). Decryption must be performed by real-time synthesis of two private keys (the algorithm is limited to Shamir). The decryption process must be audited and completed within a reasonable timeframe, ensuring data integrity.

[0053] Decrypted files are only available for viewing online by authorized personnel in a controlled environment. Output is typically restricted through video recording and screen-only viewing, prohibiting downloading or copying. The entire audit process (including decryption) is recorded in detail and an audit report is generated.

[0054] Although the above embodiments have described the concept and embodiments of the present invention in detail, those skilled in the art will recognize that various improvements and modifications can still be made to the present invention without departing from the scope of the claims, and therefore will not be elaborated here.

Claims

1. A method for graded masking and content filtering of privacy data in railway communications, characterized in that: The method includes the following steps: A main processing server is deployed at the core node of the railway communication network. A sensitive target detection model is established, verified and activated on the main processing server. The sensitive target detection model is preset with multiple types of sensitive target features. Multi-level user permission rules are configured in the background of the main processing server, and each level of user permission rules has different privacy processing principles; The main processing server receives and processes video streams and / or audio streams, wherein: When the main processing server receives the video stream, it preprocesses the video stream and extracts keyframe images. The sensitive target detection model analyzes the preprocessed keyframe image, identifies the sensitive target features contained in the keyframe image, and tracks the sensitive targets containing the sensitive target features. The main processing server obtains the user permission level of the terminal user requesting to view the video stream according to the identity authentication interface, and calls the corresponding privacy processing principle to perform privacy processing on the key frame images. The processed key frame images and non-key frame images are then recombined to form the processed video stream. When the main processing server receives the voice stream, it preprocesses the voice stream and extracts the valid voice segments. The system extracts voiceprint features from valid speech segments, converts the voiceprint features into timbre, transcribes the timbre-converted speech stream into text in real time, compares the transcribed text with a preset dynamic sensitive keyword library, and if a sensitive keyword is detected that is contained in the dynamic sensitive keyword library, the corresponding speech segment in the speech stream is replaced with a prompt sound. The processed video and / or audio streams are sent to the end users who request to view them and recorded. The original received video and audio streams are encrypted before being stored. When a security audit is required, it is performed jointly by multiple authorized personnel with specific permissions. The main processing server verifies the identity and permissions of the authorized personnel and decrypts the target file through a security mechanism for the authorized personnel to view online in a controlled environment.

2. The method for graded masking and content filtering of privacy data in railway communication according to claim 1, characterized in that: The multi-level user permission rules are divided into three levels of permissions, including: Level 1 permission: Apply Gaussian blur with a relatively large blur radius to the face area, and cover the specific ID number area with an opaque rectangle; Level 2 permissions: Apply Gaussian blur with a relatively small blur radius or edge-preserving blur to the face area, and use a semi-transparent overlay to cover specific ID number areas; Level 3 access control: Overlay a specific watermark on the face area, and fully display and add border annotations to a specific ID number area.

3. The method for graded masking and content filtering of privacy data in railway communication according to claim 1, characterized in that: When the main processing server receives the video stream, it performs keyframe image extraction processing on the video stream at certain intervals and applies image enhancement algorithms to the extracted keyframe images.

4. The method for graded masking and content filtering of privacy data in railway communication according to claim 1, characterized in that: For sensitive targets, mark their locations on keyframe images.

5. A method for graded masking and content filtering of privacy data in railway communication according to claim 1, characterized in that: For moving sensitive targets, a target tracking algorithm is enabled to predict their position in subsequent keyframe images, and re-detection is performed when the prediction deviation exceeds a threshold, ensuring that moving sensitive targets are continuously locked in the video stream.

6. A method for graded masking and content filtering of privacy data in railway communication according to claim 1, characterized in that: Use an applied image fusion algorithm to process the boundary between the occluded area and the original background.

7. A method for graded blocking and content filtering of privacy data in railway communication according to claim 1, characterized in that: When the main processing server receives the voice stream, it applies a noise reduction algorithm to remove background noise from the voice stream and retains the main human voice frequency bands; when a silent segment is detected, it is marked as invalid voice and no further processing is performed.

8. A method for graded blocking and content filtering of privacy data in railway communication according to claim 1, characterized in that: When multiple authorized personnel operate together, physical private keys are used for triggering, and decryption is achieved by synthesizing multiple physical private keys in real time.

9. A method for graded blocking and content filtering of privacy data in railway communication according to claim 1, characterized in that: The filenames of the original received video and audio streams are obfuscated, and the generated log files are encrypted and stored. These log files are periodically or in real-time synchronized to nodes or systems with auditing privileges.

10. A method for graded blocking and content filtering of privacy data in railway communication according to claim 9, characterized in that: The log files are stored on the railway department's main processing server or a dedicated storage device.