End-side AI-based video stream real-time tamper prevention review method and set-top box

Abstract

The application discloses a kind of based on end side AI's video stream real-time tamper-proofing review method and set top box, method includes: receiving and decoding video stream obtains original video frame data;Adopt adaptive frame sampling strategy to video frame is sampled;Sampling frame is input built-in AI content security review engine, and content feature vector is extracted by lightweight convolutional neural network model;The feature vector is compared with benchmark feature, and whether the content is tampered with according to whether similarity is lower than threshold value is judged, and then corresponding security response operation is triggered.The application realizes real-time intelligent review to video stream by built-in lightweight AI engine in set top box end, effectively identifies and blocks content tampering in transmission;Adopt adaptive sampling and lightweight model, while ensuring high detection rate, significantly reduce delay and resource occupation, do not affect user experience, applicable to live broadcast, on-demand and other various businesses, provide reliable and safe guarantee for the last kilometer of video distribution.

Description

Technical Field

[0001] This invention relates to the field of video security technology, specifically to a real-time anti-tampering review method for video streams based on edge AI and a set-top box. Background Technology

[0002] In traditional IPTV or digital television systems, the set-top box primarily plays a passive role in receiving and decoding video content. The front-end system is responsible for encrypting and scrambling the video content and distributing it to the terminal, while the set-top box is responsible for receiving streaming media, performing decryption and decoding operations, and ultimately presenting the video image on the user's screen. This centralized management architecture has long been the industry standard.

[0003] However, this architecture has a significant security blind spot: the video stream is vulnerable to hijacking and tampering during its transmission from the front-end server to the user's set-top box. If an attacker successfully executes a man-in-the-middle attack, replacing the original legitimate video stream with pre-prepared alternative content, the set-top box, lacking the ability to verify the authenticity of the content, will unknowingly play the altered video. This tampered content may contain inappropriate or illegal information, leading to a poor user experience and potentially causing serious broadcast security incidents, resulting in significant negative impacts on operators and content providers.

[0004] Existing security solutions primarily focus on front-end digital rights management, encryption of transmission links, or playback authentication of user identities. At the set-top box end, traditional security measures are limited to IP verification of the signal source or simple obfuscation of abnormal videos. These methods are inherently ineffective in defending against direct substitution attacks targeting the semantic layer of video content; they struggle to achieve real-time accurate identification and cannot immediately block attacks after identification. Overall, the efficiency and intelligence of protection need improvement. Summary of the Invention

[0005] This invention provides a real-time anti-tampering review method for video streams based on edge AI, which solves the security risk problem that existing set-top boxes cannot identify and block video stream tampering in the transmission link in real time at the edge.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, this application provides a real-time anti-tampering review method for video streams based on edge AI, applied to a set-top box, the method comprising:

[0008] Receive video streams from the front end and perform demultiplexing, decryption, and decoding operations to obtain the original video frame data;

[0009] An adaptive frame sampling strategy is used to sample the decoded original video frame data to obtain video frames.

[0010] The sampled video frames are input to the AI ​​content security review engine built into the set-top box, and the AI ​​content security review engine extracts the content feature vector of the video frames through a lightweight convolutional neural network model;

[0011] The content feature vector is compared with the baseline feature, and a review decision is generated based on the comparison result. The comparison includes calculating the similarity between the content feature vector and the baseline feature, and determining that the content has been tampered with when the similarity is lower than a preset threshold.

[0012] Based on the review decision, perform the appropriate security response actions on the video stream.

[0013] A further optimization scheme is that the adaptive frame sampling strategy includes:

[0014] Scene transition detection is performed on consecutive video frames based on histogram differences. When a scene transition is detected, the current video frame is forcibly sampled.

[0015] When no scene change is detected, video frames are sampled at a preset fixed frame interval.

[0016] A further optimization scheme is that the adaptive frame sampling strategy also includes:

[0017] Monitor the system load of the set-top box and dynamically adjust the fixed frame interval based on the system load.

[0018] A further optimization is that the lightweight convolutional neural network model is an improved version of MobileNetV3-Small, quantized using INT8.

[0019] A further optimization scheme is that the benchmark features include dynamic benchmarks and static benchmarks;

[0020] The dynamic benchmark is a content fingerprint corresponding to the currently playing program, which is synchronously sent by the front end through a private protocol.

[0021] The static baseline is a pre-configured security policy model corresponding to a specific channel or program type.

[0022] A further optimization scheme is that the security response operation includes at least one of interrupting playback, replacing the playback screen with a warning message or log reporting.

[0023] Secondly, this application provides a set-top box, comprising:

[0024] The receiving and decoding module is used to receive the video stream from the front end and perform demultiplexing, decryption, and decoding operations to obtain the original video frame data.

[0025] The sampling module is communicatively connected to the receiving and decoding module and is used to sample the decoded original video frame data to obtain video frames using an adaptive frame sampling strategy.

[0026] The AI ​​content security review engine is connected to the sampling module and is used to receive video frames from the sampling module, extract the content feature vector of the video frame through a lightweight convolutional neural network model, and compare the content feature vector with the benchmark feature to generate a review decision.

[0027] The security response module is communicatively connected to the AI ​​content security review engine and the receiving and decoding module, and is used to perform corresponding security response operations on the video stream based on the review decision.

[0028] A further optimization is that the AI ​​content security review engine is integrated into the main system-on-a-chip of the set-top box, and it takes the form of a hardware neural network processing unit or a high-priority software thread.

[0029] A further optimization is that the benchmark features include dynamic benchmarks, which are synchronously distributed from the front end to the AI ​​content security review engine via a private protocol;

[0030] The AI ​​content security review engine is configured to calculate the cosine similarity between the content feature vector extracted in real time and the dynamic benchmark to generate review decisions.

[0031] Thirdly, this application provides a computer-readable storage medium storing a real-time anti-tampering review program for video streams based on edge AI, wherein when the real-time anti-tampering review program for video streams based on edge AI is executed by a processor, it implements the steps of the real-time anti-tampering review method for video streams based on edge AI as described above.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0033] By embedding a lightweight AI content security review engine into the set-top box decoding and playback link, real-time intelligent review of playback content is achieved, fundamentally improving end-side security. It can effectively identify and block illegal insertions and content tampering behaviors such as replacement of harmful information caused by man-in-the-middle attacks, thereby preventing security broadcast incidents.

[0034] By designing an adaptive frame sampling strategy and a dedicated lightweight convolutional neural network model, while ensuring a high detection rate for both known and unknown tampering types, performance and resource consumption are greatly optimized, achieving extremely low end-to-end review latency, ensuring no impact on the user's viewing experience, and making it suitable for various terminal environments with limited resources.

[0035] It can seamlessly cover various business scenarios such as live streaming and video-on-demand, providing a reliable security barrier for the last mile of video distribution. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0037] Figure 1 A flowchart of a real-time anti-tampering review method for video streams based on edge AI provided in this application embodiment;

[0038] Figure 2 This is a functional block diagram of a real-time anti-tampering review system for video streams based on edge AI, provided in an embodiment of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0040] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0041] INT8: 8-bit Integer Quantization;

[0042] AI: Artificial Intelligence;

[0043] ReLU6: Rectified Linear Unit 6;

[0044] NPU: Neural Processing Unit;

[0045] SoC: System on Chip;

[0046] IPTV: Internet Protocol Television.

[0047] IP: Internet Protocol.

[0048] MPEG-2 TS: Moving Picture Experts Group-2 Transport Stream;

[0049] DRM: Digital Rights Management.

[0050] RGB: Red, Green, Blue;

[0051] I / O: Input / Output.

[0052] Firstly, such as Figure 1 As shown, this application provides a real-time anti-tampering review method for video streams based on edge AI, applied to a set-top box. The method includes:

[0053] Step S1: Receive the video stream from the front end and perform demultiplexing, decryption, and decoding operations to obtain the original video frame data;

[0054] Step S2: Using an adaptive frame sampling strategy, sample the decoded original video frame data to obtain video frames;

[0055] Step S3: Input the sampled video frames into the AI ​​content security review engine built into the set-top box, and the AI ​​content security review engine extracts the content feature vector of the video frames through a lightweight convolutional neural network model;

[0056] Step S4: Compare the content feature vector with the baseline feature, and generate a review decision based on the comparison result. The comparison includes calculating the similarity between the content feature vector and the baseline feature, and determining that the content has been tampered with when the similarity is lower than a preset threshold.

[0057] Step S5: Perform the appropriate security response operation on the video stream based on the review decision.

[0058] This embodiment integrates a lightweight AI content security review engine into the set-top box decoding and playback chain, enabling real-time intelligent review of playback content. Its core effect lies in fundamentally improving end-side security, effectively identifying and blocking illegal insertions and content tampering caused by man-in-the-middle attacks, thus preventing broadcast security incidents. Through an adaptive frame sampling strategy and a dedicated lightweight convolutional neural network model, the solution ensures a high detection rate for both known and unknown tampering types while significantly optimizing performance and resource consumption, achieving extremely low end-to-end review latency. This ensures no impact on the user's viewing experience and is applicable to various resource-constrained terminal environments. Furthermore, this solution has broad applicability, seamlessly covering various business scenarios such as live streaming and video-on-demand, providing a reliable security barrier for the last mile of video distribution.

[0059] In one embodiment, step S1: receiving and decoding the video stream from the front end, converting it into processable raw video frame data, includes the following steps:

[0060] Step S11: Perform demultiplexing on the encrypted video transport stream from the front end, separating it into independent video and audio elementary streams;

[0061] The front end is the source system responsible for video content production, encryption, scrambling, distribution, and generating content fingerprints;

[0062] Specifically, demultiplexing usually refers to breaking down and reassembling a received composite transport stream (such as an MPEG-2 TS stream) into independent video and audio elementary streams based on the program association table and data packet identifiers contained therein, for subsequent processing.

[0063] Step S12: Perform a decryption operation on the separated video elementary stream to restore the plaintext video compressed data;

[0064] Specifically, the encrypted video base stream data is decrypted using the decryption algorithm and key corresponding to the front-end encryption (usually provided by a digital rights management or conditional access system), restoring it to a standard, unencrypted video compressed stream.

[0065] Step S13: Perform a decoding operation on the decrypted video compressed data to convert it into raw video frame data in YUV or RGB format;

[0066] Specifically, the compressed bitstream is decoded by a video decoder according to the corresponding video encoding standard, and the compressed data is restored into a series of complete, uncompressed original video frame images, whose pixel format is usually YUV or RGB, so that they can be directly used for display or further image processing.

[0067] This embodiment accurately separates the video elementary stream from the composite stream through demultiplexing. The decryption operation ensures the security of the transmitted content and restores the readability of the data. Finally, the compressed video data is converted into raw pixel frames in standard formats such as YUV or RGB through decoding. This ensures that the video frame data input to the AI ​​engine has both content integrity and format uniformity, thus laying a solid data foundation for subsequent high-precision feature extraction and comparison.

[0068] In one embodiment, step S2: using an adaptive frame sampling strategy to sample the decoded original video frame data to obtain video frames, including the following steps:

[0069] Step S21: Calculate the histogram distribution difference between the current video frame and the previous video frame in the YUV color space;

[0070] Step S22: Compare the calculated histogram difference value with a preset threshold to determine whether a scene switch has occurred;

[0071] Step S23: If the difference value exceeds the preset threshold, it is determined to be a scene switch, and the current video frame is forcibly selected as the sampling frame;

[0072] Specifically, the preset threshold used to determine scene switching can be set to 30%; at the same time, when multiple consecutive scene switching is detected, the sampling frequency will be temporarily increased to ensure that key content is covered; specifically, the temporarily increased sampling frequency can be adjusted to sample once every 10 frames.

[0073] Step S24: If no scene change is detected, then select video frames from the video frame sequence regularly as sampling frames according to a preset fixed frame interval.

[0074] Specifically, the initial value of the fixed frame interval can be configured to sample once every 30 frames, which corresponds to a video stream of 25 or 30 frames per second, achieving a review frequency of approximately once per second. The system also monitors the real-time system load of the set-top box in real time and dynamically adjusts the fixed sampling interval according to the load to balance review coverage and system performance. Specifically, when the system load is detected to exceed 80%, the sampling interval is automatically extended. More specifically, the extended interval can be adjusted to sample once every 60 frames.

[0075] This embodiment accurately captures scene transition events by calculating inter-frame differences in real time and comparing them with thresholds. This ensures that review is triggered immediately at critical moments when video content changes significantly, thus avoiding the omission of tampering caused by sudden scene changes. During stable periods, regular sampling is performed at fixed intervals, providing a basic guarantee for continuous security monitoring. More importantly, this strategy introduces a dynamic adjustment mechanism based on system load. This mechanism allows the system to automatically reduce the sampling frequency to alleviate processing burden when device computing resources are strained, and temporarily increase the sampling density during periods of high information volume with frequent content changes. This adaptive capability enables the entire review system to intelligently and dynamically optimize the use of computing resources while ensuring high coverage review of key frames in the video stream. Ultimately, this achieves the goal of efficient and accurate real-time security protection without affecting the user's viewing smoothness or the normal performance of the terminal device.

[0076] In one embodiment, step S3: The sampled video frame is input to the AI ​​content security review engine built into the set-top box. The AI ​​content security review engine extracts the content feature vector of the video frame through a lightweight convolutional neural network model, including the following steps:

[0077] Step S31: Input the selected video frame data into the set-top box's built-in AI content security review engine;

[0078] The AI ​​content security review engine can be implemented as an independent hardware unit integrated into the main system-on-a-chip of the set-top box, or as a high-priority software processing thread.

[0079] Step S32: Call and run the lightweight convolutional neural network model to perform inference calculations on the video frames;

[0080] Specifically, the model is based on the improved MobileNetV3-Small architecture and contains a total of 15 computational layers, consisting of three 3x3 standard convolutional layers, nine depthwise separable convolutional blocks, two fully connected layers, and one global average pooling layer. The model uses the Hard-Swish activation function in the deep network and the ReLU6 activation function in the shallow network to balance computational accuracy and efficiency. After a series of convolutional and nonlinear feature transformations, the model outputs a 128-dimensional normalized feature vector from its last fully connected layer, which serves as the content feature vector representing the deep semantics of the video frame.

[0081] Preferably, the model has undergone INT8 quantization, which compresses its storage size to approximately 1.2MB. The time required to complete a single inference on the terminal NPU can be less than 10 milliseconds, thus efficiently adapting to the computing and storage resource limitations of terminal devices. In addition, the model was trained on a large-scale dataset in the cloud before deployment. The training set contains more than 100,000 normal video frames and more than 50,000 tampered samples covering various types, including malicious images, illegal logos, unauthorized advertisements, and image replacements. This gives it powerful feature generalization and discrimination capabilities.

[0082] Step S33: Obtain the normalized content feature vector representing the semantics of the video from the output layer of the lightweight convolutional neural network model;

[0083] Specifically, the lightweight convolutional neural network model processes the output of its last fully connected layer to generate a vector with a fixed 128 dimensions and normalization. This vector serves as the content feature vector that ultimately represents the deep semantic content of the video frame, providing core data for subsequent feature comparison.

[0084] This embodiment successfully achieved efficient semantic feature extraction on resource-constrained set-top box terminals by deploying a lightweight convolutional neural network model that has undergone deep optimization and INT8 quantization. The model is only about 1.2MB in size, with a single inference latency of less than 10 milliseconds on the terminal's NPU, resulting in extremely low system resource consumption. At the same time, the model is trained on a large-scale dataset containing over 100,000 normal frames and 50,000 tampered samples, exhibiting strong generalization ability and stably outputting a highly discriminative 128-dimensional normalized content feature vector. This provides reliable core data for subsequent real-time and accurate content security comparison decisions and is a key step in applying deep learning capabilities to real-time video stream review on terminals.

[0085] In one embodiment, step S4: comparing the extracted real-time content feature vector with the baseline features based on similarity, and generating a review decision accordingly, includes the following steps:

[0086] Step S41: Obtain the baseline features;

[0087] Specifically, the baseline features are obtained through the following two methods:

[0088] One type is a dynamic benchmark, which is synchronously distributed by the front-end system through a private protocol. Its essence is that the front-end AI generates a 128-dimensional content fingerprint of the original program master tape in real time and corresponds to the currently playing program.

[0089] Another type is the static baseline, which is a security policy model feature that is pre-configured and embedded inside the set-top box and corresponds to a specific channel or program type;

[0090] Specifically, dynamic benchmarks are used to perform precise verification using the original program as the gold standard, such as verifying the consistency of live streams; static benchmarks are used to make conformity judgments based on semantic categories, such as determining whether game scenes that should not be present appear in news channel footage.

[0091] Step S42: Calculate the content similarity between the real-time features and the baseline features;

[0092] Specifically, the cosine similarity algorithm is used to perform the calculation, which quantifies the directional consistency between two high-dimensional vectors into a scalar value within a specific range;

[0093] Step S43: Compare the calculated similarity value with the preset decision threshold to obtain the threshold comparison result;

[0094] Step S44: Output the final review decision based on the threshold comparison results;

[0095] Specifically, if the similarity is lower than the preset judgment threshold, it is determined that the current video content has been tampered with or poses a security risk, and a decision to reject the review is output.

[0096] If the similarity is equal to or higher than the threshold, the content is deemed safe and a decision to pass the review is output.

[0097] This embodiment constructs a comparison system that is both accurate and semantically generalizable by integrating dynamic real-time fingerprints and static policy models as dual benchmarks. It uses cosine similarity to transform abstract semantic consistency into quantifiable and comparable values, and generates clear and reliable binary security decisions by comparing them with explicit thresholds. This process enables the set-top box to independently complete complex content security judgments on the device side, providing a core and reliable basis for triggering immediate and automated security responses.

[0098] In one embodiment, step S5: Performing appropriate security response operations on the video stream based on the review decision includes the following steps:

[0099] Step S51: If the review decision is passed, the corresponding video frame data is input normally into the display buffer and handed over to the display device for rendering and playback;

[0100] Step S52: If the review decision is not approved, a control command is immediately sent to the video decoding module to forcibly interrupt the decoding and playback process of the current video stream. This interruption operation can be specifically manifested as a black screen or mute, thereby immediately blocking the spread of the tampered content. At the same time as interrupting playback, the control display device replaces the current playback screen with a pre-set warning information interface to intuitively prompt the user about the current situation.

[0101] Specifically, the displayed warning message may include explicit prompts such as "Content security review failed, suspected signal source has been tampered with";

[0102] Step S53: Encapsulate the information related to this tampering alarm event, including the event type, key feature vectors or screenshots of the tampered content, and precise timestamps, into a standard security log, and report it to the front-end management system for operation and maintenance personnel to audit, trace, and conduct subsequent investigations and handling.

[0103] This embodiment transforms the intelligent analysis and decision-making of the preceding steps into immediately visible and controllable security operations on the terminal device through a clear and rapid execution closed loop. It ensures that secure content passes through unnoticed, and once a risk of tampering is identified, it can immediately block the harm, alert the user, and report the incident. Thus, it constitutes a complete and effective terminal security protection response at three levels: user experience, security blocking, and operation and maintenance traceability.

[0104] Based on simulations of typical 4K set-top box hardware specifications (main SoC 1.8GHz, NPU computing power 1.0 TOPS), the expected performance indicators of this solution are shown in Table 1:

[0105] Table 1 System Design Objectives and Expected Performance Indicators

[0106]

[0107] The above indicators show that, in theory, this solution can achieve high accuracy and low latency real-time video tampering review with extremely low resource overhead, and further verification can be achieved through actual testing.

[0108] Secondly, such as Figure 2 As shown, this application provides a set-top box, including a housing and a receiving decoding module 100, a sampling module 200, an AI content security review engine 300 and a security response module 400 disposed within the housing;

[0109] The receiving and decoding module 100 is used to receive the video stream from the front end and perform demultiplexing, decryption and decoding operations to obtain the original video frame data;

[0110] The sampling module 200 is communicatively connected to the receiving and decoding module 100 and is used to sample the decoded original video frame data to obtain video frames using an adaptive frame sampling strategy.

[0111] The AI ​​content security review engine 300 is communicatively connected to the sampling module 200, and is used to receive video frames from the sampling module, extract the content feature vector of the video frame through a lightweight convolutional neural network model, and compare the content feature vector with the benchmark feature to generate a review decision.

[0112] The security response module 400 is communicatively connected to the AI ​​content security review engine 300 and the receiving and decoding module 100, and is used to perform corresponding security response operations on the video stream according to the review decision.

[0113] In one embodiment, the AI ​​content security review engine is integrated into the main system-on-a-chip of the set-top box, and takes the form of a hardware neural network processing unit or a high-priority software thread.

[0114] In one embodiment, the benchmark feature includes a dynamic benchmark, which is synchronously distributed from the front end to the AI ​​content security review engine via a private protocol;

[0115] The AI ​​content security review engine is configured to calculate the cosine similarity between the content feature vector extracted in real time and the dynamic benchmark to generate review decisions.

[0116] The functions of each module in the above-mentioned AI-based real-time anti-tampering review system for video streams correspond to the steps in the above-mentioned AI-based real-time anti-tampering review method for video streams. Their functions and implementation processes will not be described in detail here.

[0117] Thirdly, this application provides a real-time anti-tampering review device for video streams based on edge AI. Such a device can be a personal computer (PC), laptop, server, or other device with data processing capabilities.

[0118] In this application embodiment, a real-time anti-tampering review device for video streams based on edge AI may include a processor, a memory, a communication interface, and a communication bus.

[0119] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0120] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting internal components of a real-time anti-tampering review device for video streams based on edge AI, as well as interfaces for interconnecting the device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0121] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0122] The processor can be a general-purpose processor, which can call a real-time anti-tampering review program for video streams based on edge AI stored in memory and execute the real-time anti-tampering review method for video streams based on edge AI provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the real-time anti-tampering review program for video streams based on edge AI is called can refer to the various embodiments of the real-time anti-tampering review method for video streams based on edge AI of this application, and will not be repeated here.

[0123] Fourthly, embodiments of this application also provide a readable storage medium.

[0124] This application stores a real-time anti-tampering review program for video streams based on edge AI on a readable storage medium. When the real-time anti-tampering review program for video streams based on edge AI is executed by a processor, it implements the steps of the real-time anti-tampering review method for video streams based on edge AI as described above.

[0125] One of the methods implemented when a real-time anti-tampering review procedure for video streams based on edge AI is executed can be referred to in various embodiments of the real-time anti-tampering review method for video streams based on edge AI in this application, and will not be repeated here.

[0126] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time anti-tampering review method for video streams based on edge AI, characterized in that, Applied to a set-top box, the method includes: Receive video streams from the front end and perform demultiplexing, decryption, and decoding operations to obtain the original video frame data; An adaptive frame sampling strategy is used to sample the decoded original video frame data to obtain video frames. The sampled video frames are input to the AI ​​content security review engine built into the set-top box, and the AI ​​content security review engine extracts the content feature vector of the video frames through a lightweight convolutional neural network model; The content feature vector is compared with the baseline feature, and a review decision is generated based on the comparison result. The comparison includes calculating the similarity between the content feature vector and the baseline feature, and determining that the content has been tampered with when the similarity is lower than a preset threshold. Based on the review decision, perform the appropriate security response actions on the video stream.

2. The real-time anti-tampering review method for video streams based on edge AI according to claim 1, characterized in that, The adaptive frame sampling strategy includes: Scene switching detection based on histogram differences is performed on consecutive video frames; When a scene change is detected, sampling of the current video frame is forced. When no scene change is detected, video frames are sampled at a preset fixed frame interval.

3. The real-time anti-tampering review method for video streams based on edge AI according to claim 2, characterized in that, The adaptive frame sampling strategy also includes: Monitor the system load of the set-top box and dynamically adjust the fixed frame interval based on the system load.

4. The real-time anti-tampering review method for video streams based on edge AI according to claim 1, characterized in that, The lightweight convolutional neural network model is an improved version of MobileNetV3-Small, quantized using INT8.

5. The real-time anti-tampering review method for video streams based on edge AI according to claim 1, characterized in that, The reference features include dynamic references and static references; The dynamic benchmark is a content fingerprint corresponding to the currently playing program, which is synchronously sent by the front end through a private protocol. The static baseline is a pre-configured security policy model corresponding to a specific channel or program type.

6. The real-time anti-tampering review method for video streams based on edge AI according to claim 1, characterized in that, The security response operation includes at least one of interrupting playback, replacing the playback screen with a warning message or log report.

7. A set-top box, characterized in that, include: The receiving and decoding module is used to receive the video stream from the front end and perform demultiplexing, decryption, and decoding operations to obtain the original video frame data. The sampling module is communicatively connected to the receiving and decoding module and is used to sample the decoded original video frame data to obtain video frames using an adaptive frame sampling strategy. The AI ​​content security review engine is connected to the sampling module and is used to receive video frames from the sampling module, extract the content feature vector of the video frame through a lightweight convolutional neural network model, and compare the content feature vector with the benchmark feature to generate a review decision. The security response module is communicatively connected to the AI ​​content security review engine and the receiving and decoding module, and is used to perform corresponding security response operations on the video stream based on the review decision.

8. The set-top box according to claim 7, characterized in that, The AI ​​content security review engine is integrated into the main system-on-a-chip of the set-top box, and it takes the form of a hardware neural network processing unit or a high-priority software thread.

9. The set-top box according to claim 7, characterized in that, The benchmark features include dynamic benchmarks, which are synchronously distributed from the front end to the AI ​​content security review engine via a private protocol. The AI ​​content security review engine is configured to calculate the cosine similarity between the content feature vector extracted in real time and the dynamic benchmark to generate review decisions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a real-time anti-tampering review program for video streams based on edge AI, wherein when the real-time anti-tampering review program for video streams based on edge AI is executed by a processor, it implements the steps of a real-time anti-tampering review method for video streams based on edge AI as described in any one of claims 1 to 6.

Images