A video playing control method, a computer device and a storage medium

Abstract

Embodiments of the present disclosure provide a video playback control method, a computer device and a storage medium, to bind the audit conclusion with the playback content through watermarking, and to verify the watermark load and malicious editing in real time at the terminal, so as to realize a playback security closed loop. The video playback control method comprises: obtaining a watermarked video file distributed to the local, the watermarked video file being generated and distributed by a server calling a neural network watermark embedder to embed a digital watermark in an audited video file; the neural network watermark embedder is jointly trained with a neural network watermark verifier on the side of a broadcast control terminal; in a video playback processing link, the neural network watermark verifier is called to verify the playback frame, and a verification result is obtained; the neural network watermark verifier comprises a watermark load recovery unit and a malicious editing detection unit; if the verification result indicates that the watermark check is passed and no malicious editing is detected, the playback is continued; otherwise, a broadcast control blocking action is performed and failure information is reported.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a video playback control method, computer device, and storage medium. Background Technology

[0002] In video content distribution and broadcast control systems, content security and compliance review are crucial for ensuring the legality and compliance of broadcast content. Currently, mainstream content security solutions typically employ an architecture that separates service center review from terminal playback.

[0003] Common approaches in related technologies include: direct distribution and playback after review by the service center, relying on post-event random checks to identify problems; and solutions that combine Digital Rights Management (DRM) or encrypted licensing to control playback permissions. Additionally, there are solutions that use watermarking technology to mark video content for source tracing or detection.

[0004] However, in the method of directly distributing content after service center review, there is a lack of effective verification mechanism between the service center review and the terminal playback, making it difficult to ensure that the content actually played on the terminal is the original content that has passed the review. In addition, the relevant watermarking technology is only used for tracing or detection, and lacks the real-time judgment and blocking capabilities that are deeply integrated with the playback link on the terminal side, making it difficult to detect and deal with anomalies such as content tampering in a timely manner. Summary of the Invention

[0005] This disclosure provides a video playback control method that binds the review conclusion to the playback content through a watermark and verifies the watermark payload and malicious editing in real time on the terminal to achieve a closed loop of playback security.

[0006] Firstly, a video playback control method is provided, applied to a broadcast control terminal. The method includes: acquiring a watermarked video file distributed locally, wherein the watermarked video file is generated and distributed by a server calling a neural network watermark embedder to embed a digital watermark into an approved video file; the neural network watermark embedder is obtained by generating training video data based on benign transformations and malicious editing through dynamic data augmentation, and then co-training it with a neural network watermark verifier on the broadcast control terminal side by optimizing a joint loss function; wherein the watermark embedded by the neural network watermark embedder is invisible to the naked eye and robust to preset benign transformations, and can be verified by the neural network watermark verifier. Successful recovery demonstrates sensitivity to malicious editing, enabling the neural network watermark verifier to detect anomalies. In the video playback processing chain, the neural network watermark verifier is invoked to verify the playback frame, yielding a verification result. The neural network watermark verifier includes a watermark payload recovery unit and a malicious editing detection unit. The watermark payload recovery unit recovers the watermark payload from the playback frame, and the malicious editing detection unit detects whether the playback frame has been maliciously edited. The verification result includes a watermark verification result and a malicious editing detection result. If the verification result indicates successful verification, video playback continues. If the verification result indicates unsuccessful verification, a broadcast control blocking action is executed, and a failure message is reported.

[0007] In one implementation, the step of calling the neural network watermark verifier to verify the playback frame in the video playback processing link includes: calling the neural network watermark verifier frame by frame for verification after video frame decoding and before rendering and display; or, sampling one frame every preset number of frames and calling the neural network watermark verifier for verification.

[0008] In one implementation, the step of calling the neural network watermark verifier to verify the playback frame and obtain the verification result includes: inputting the playback frame into the neural network watermark verifier to obtain the watermark payload output by the watermark payload recovery unit and the tampering heatmap or tampering score output by the malicious editing detection unit; if the tampering score is greater than a preset tampering score threshold, or the maximum value of the tampering heatmap is greater than a preset maximum value threshold, or the average value of the tampering heatmap is greater than a preset average value threshold, then malicious editing is determined to be detected; if the confidence level, bit error rate, or p-value index of the watermark payload does not meet the corresponding threshold, then watermark verification is determined to be unsuccessful.

[0009] In one implementation, if the verification result indicates that the verification failed, a broadcast control blocking action is executed and a failure message is reported, including: obtaining the verification results of multiple playback frames within a preset time window, and counting the number of frames that were determined to be maliciously edited or the number of frames that failed watermark verification within the preset time window; if the count result exceeds a preset threshold, and it is determined that the verification result indicates that the verification failed, then the broadcast control blocking action is triggered and a failure message is reported.

[0010] In one implementation, the step of executing the broadcast control blocking action and reporting failure information includes: executing the broadcast control blocking action, which includes at least one of pausing playback, displaying a black screen, cutting off the video output signal, or switching to backup video content; synchronously extracting the content identifier, timestamp, failure type, frame summary information, and terminal device identifier of the playback frame, generating failure information, and reporting it to the server; wherein, the failure type includes watermark verification failure and / or detection of malicious editing.

[0011] In one implementation, the neural network watermarking validator and the server-side neural network watermarking embedder are jointly trained as follows: Training video data is acquired, and training enhancements are divided into a benign transformation set and a malicious editing set; the benign transformation set includes video encoding, video transcoding, image quality degradation, color or brightness adjustment, and minor geometric transformations; the malicious editing set includes text rendering overlay, image overlay, local patching, and cropping or trimming; transformations from the benign transformation set are applied to the training video data to obtain benign transformation samples; transformations from the malicious editing set are applied to the training video data, and pixel-level mask labels are generated to obtain malicious editing samples and their corresponding pixel-level mask labels; based on the benign transformation samples, the malicious editing samples, and their corresponding pixel-level mask labels, the neural network watermarking validator and the neural network watermarking embedder are jointly trained.

[0012] In one implementation, during joint training, the neural network watermark embedder is used to embed binary watermark messages into video frames to generate watermarked samples; the neural network watermark verifier is used to recover the watermark payload from the watermarked samples and detect malicious editing; the joint training of the neural network watermark verifier and the neural network watermark verifier based on the benign transformation samples and the maliciously edited samples and their corresponding pixel-level mask labels includes: calculating the watermark payload recovery loss for the neural network watermark verifier and the neural network watermark embedder, and the imperceptible loss for the neural network watermark embedder, based on the benign transformation samples. The system includes a constraint loss and a false positive suppression loss for the neural network watermark verifier. The false positive suppression loss is used to constrain the tampering heatmap output by the malicious editing detection unit under benign transformation to be close to all zeros. Based on the malicious editing sample and the corresponding pixel-level mask label, the tampering detection loss for the malicious editing detection unit of the neural network watermark verifier is calculated. The watermark payload recovery loss, the imperceptibility constraint loss, the false positive suppression loss and the tampering detection loss are weighted and summed to obtain a joint loss function. The neural network watermark verifier and the neural network watermark embedder are trained collaboratively by optimizing the joint loss function.

[0013] In one implementation, the neural network watermark verifier and the neural network watermark embedder are jointly trained by optimizing the joint loss function, including: performing a first-stage training: using image data as training samples, jointly training the neural network watermark embedder and the neural network watermark verifier, and optimizing the watermark payload recovery loss and imperceptible constraint loss in the joint loss function; performing a second-stage training: using image data and video data as training samples, training with benign transformations related to video compression and time watermark propagation strategies, and training with maliciously edited samples and their pixel-level mask labels, optimizing the tamper detection loss and false alarm suppression loss in the joint loss function; and performing a third-stage training: freezing the network parameters of the neural network watermark embedder, fine-tuning the neural network watermark verifier, and optimizing the tamper detection loss and false alarm suppression loss in the joint loss function.

[0014] Secondly, this disclosure also provides a video playback control method applied to a server. The method includes: acquiring a video file to be published; performing content review on the video file to be published; and outputting a review conclusion; if the review conclusion indicates that the review is passed, calling a neural network watermark embedder to embed a neural network digital watermark in the frames of the video file to be published to generate a watermarked video file; and distributing the watermarked video file to a broadcast control terminal for the broadcast control terminal to execute the aforementioned video playback control method.

[0015] In one embodiment, embedding a neural network digital watermark in the frames of the video file to be published includes: in response to the review conclusion indicating that the review has been passed, extracting the content identifier, policy version number, and review time from the review conclusion and encoding them into a binary message; embedding the binary message as watermark information into the frames of the video file to be published, so that the watermarked video file carries the unique identifier of the review conclusion.

[0016] In one embodiment, embedding a neural network digital watermark in the frames of the video file to be published includes: scaling the frames of the video file to be published to a preset resolution to obtain scaled frames; calling a neural network watermark embedder to embed binary messages in the pixel domain of the scaled frames to generate watermark distortion; upsampling the watermark distortion back to the original resolution and superimposing it onto the original frame with a preset intensity coefficient to obtain a watermarked video file.

[0017] In one embodiment, embedding a neural network digital watermark in the frames of the video file to be published further includes employing a time watermark propagation strategy: performing a watermark embedding operation on one frame in every k frames to generate watermark distortion; and propagating the watermark distortion to the adjacent k-1 frames.

[0018] Thirdly, this disclosure also provides a video playback control device, applied to a broadcast control terminal, comprising:

[0019] The acquisition module is used to acquire watermarked video files distributed locally. The watermarked video files are generated and distributed by the server after embedding digital watermarks into approved video files using a neural network watermark embedder. The neural network watermark embedder is obtained by generating training video data based on benign transformations and malicious editing through dynamic data augmentation, and then co-training it with the neural network watermark verifier on the broadcast control terminal side by optimizing the joint loss function. The watermark embedded by the neural network watermark embedder is invisible to the naked eye, robust to preset benign transformations and can be successfully recovered by the neural network watermark verifier, and sensitive to malicious editing and can be detected by the neural network watermark verifier. The verification module is used to call a neural network watermark verifier to verify the playback frame in the video playback processing link and obtain the verification result. The neural network watermark verifier includes a watermark payload recovery unit and a malicious editing detection unit. The watermark payload recovery unit is used to recover the watermark payload from the playback frame, and the malicious editing detection unit is used to detect whether the playback frame has been maliciously edited. The verification result includes the watermark verification result and the malicious editing detection result. The broadcast control module is used to continue video playback if the verification result indicates that the verification is successful; and to execute a broadcast control blocking action and report failure information if the verification result indicates that the verification is unsuccessful.

[0020] Fourthly, embodiments of this disclosure also provide a video playback control device, applied to a server, comprising: The acquisition module is used to acquire the video file to be published, perform content review on the video file to be published, and output the review conclusion. The watermark embedding module is used to call the neural network watermark embedder to embed a neural network digital watermark in the frames of the video file to be published if the review conclusion indicates that the review is passed, thereby generating a watermarked video file. The sending module is used to distribute the watermarked video file to the broadcast control terminal so that the broadcast control terminal can execute the video playback control method described in any of the above embodiments.

[0021] Fifthly, embodiments of this disclosure also provide a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions to perform the video playback control method as described in any of the above embodiments.

[0022] Sixthly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the video playback control method as described in any of the above embodiments.

[0023] The video playback control method provided by the embodiments of the present disclosure effectively solves the problems of脱节 between auditing and playback and the lack of real-time security protection capabilities at the terminal by integrating a neural network watermark verifier with dual functions of watermark payload recovery and malicious editing detection into the playback link of the playback control terminal. The method first obtains a neural network watermark embedder jointly trained with the neural network watermark verifier by the server, embeds a watermark into the video file with the watermark distributed after the audit is passed, so as to bind the audit conclusion with the video content. In particular, the neural network watermark embedder is co-trained with the verifier on the terminal side and uses dynamic data augmentation to distinguish between benign transformations and malicious edits, so that the embedded watermark remains robust in the case of benign transformations in the normal distribution scenario to ensure recoverability, and at the same time is sensitive to malicious edits such as text covering and image overlay to trigger anomaly detection, so as to effectively identify tampered content on the premise of ensuring a normal playback experience. During the playback process of the playback control terminal, the neural network watermark verifier is called in real time to perform dual verification on the playback frames: verify whether the playback content is the original version passed by the audit through watermark payload recovery, and identify whether there are tampering behaviors such as text covering, image overlay, and cutting through malicious edit detection. Compared with the related technologies that only rely on service center auditing or permission control, this method extends the verification to the terminal playback link and realizes an end-to-end content verifiable closed loop. In addition, when the verification is passed, normal playback is performed, otherwise playback control is blocked and a failure message is reported. Compared with the defect of existing watermark technologies that are only used for retrospective tracing after the event and lack linkage capabilities, this method realizes real-time judgment and immediate blocking on the terminal side, can intercept abnormal content before it is broadcast, and significantly improves the security and compliance of the playback control content.

[0024] For the implementation manners and beneficial effects of the above video playback control device, computer device and storage medium, refer to the description of the method content, which will not be elaborated here.

[0025] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the technical solutions of the present disclosure.

[0026] To make the above objects, features and advantages of the present disclosure more obvious and understandable, the following specific embodiments are given and described in detail in conjunction with the accompanying drawings. Description of the Drawings

[0027] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall architecture of the distributed video broadcast control content security system provided in the embodiments of this disclosure; Figure 2 This disclosure provides a flowchart of a video playback control method according to an embodiment; Figure 3 A schematic diagram of a dual-head decoding network structure with malicious editing detection provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of a video playback control method provided in an embodiment of the present disclosure; Figure 5 This is a schematic diagram of the neural network watermark embedding and extraction process provided in the embodiments of this disclosure; Figure 6 A schematic diagram of a video playback control device 600 provided in an embodiment of this disclosure; Figure 7 A schematic diagram of another video playback control device 700 provided in an embodiment of this disclosure; Figure 8 This is a schematic diagram of the structure of a computer device 800 as an exemplary embodiment of the present disclosure. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0030] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0031] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0032] Furthermore, the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein.

[0033] Research has revealed several key aspects of video content distribution and broadcast control systems. One approach combines service center review with post-distribution sampling or manual inspection. The service center reviews the video content, and once approved, it's directly distributed to terminals for playback. Subsequent issues are identified through manual sampling or post-distribution accountability mechanisms. This method lacks the ability to continuously verify the content after approval, making it difficult to prevent content replacement or tampering during distribution or playback. Another approach combines service center review with Digital Rights Management (DRM) / encrypted authorization. DRM technology controls playback permissions, ensuring only authorized terminals can play the content. However, DRM primarily controls playback permissions and cannot verify whether the content itself is the original approved content. It also struggles to handle content tampering in scenarios involving screen recording, transcoding, splicing, or other altered distribution methods. Finally, a third approach uses watermarking technology to mark video content for traceability or detection. These watermarking schemes are mostly designed with robustness as their core objective, aiming for the recoverability of watermark information after common processing such as compression, transcoding, scaling, and filtering. These solutions typically treat watermarks as invisible labels for post-event source tracing, copyright declarations, or piracy tracking. To achieve high robustness, watermarks are often embedded in the transform domain or features with strong statistical stability, making them insensitive to semantic changes in video content. In broadcast control scenarios, they are usually applied offline or by sampling, i.e., playing first and then detecting, or periodically sampling for watermark extraction and verification. There is a lack of real-time linkage between watermark detection results and broadcast control actions; even if watermark loss or tampering is detected, it is difficult to intervene before abnormal content is broadcast. Furthermore, existing watermarking solutions treat benign transformations such as compression and transcoding as the main sources of interference, while considering malicious editing such as text overlay, image overlay, and partial cropping as processing operations of the same nature as benign transformations. Therefore, when illegal text or badges are overlaid on a video, the watermark may still be successfully recovered.

[0034] Based on this, the present disclosure provides a video playback control method, a computer device, and a storage medium, which can effectively overcome the defects of the above solutions.

[0035] First, compared to related technologies that employ service center review and subsequent random checks or manual inspections, this embodiment embeds digital watermarks into approved videos through a service center, thus binding the review conclusion to the video content. The broadcast control terminal calls a neural network watermark verifier in real time during the playback process to verify the watermark; playback only continues when the watermark verification passes and no malicious editing is detected. This mechanism extends the review and verification process to the terminal playback stage, enabling real-time verification of the playback content and effectively compensating for the shortcomings of subsequent random checks that cannot promptly detect problems.

[0036] Secondly, compared to related technologies that employ service center review and DRM / encryption authorization schemes, the neural network watermark verifier of this disclosure includes a watermark payload recovery unit and a malicious editing detection unit. The watermark payload recovery unit verifies whether the playback frame contains the watermark information of the reviewed version, and the malicious editing detection unit detects whether there are tampering behaviors such as text overlay, image superposition, or cropping. Compared to DRM, which only controls whether playback is allowed, the solution of this disclosure can verify what content is being played and identify tampering behaviors. Even if the content has undergone transformation processing such as transcoding or screen recording, the compliance of the content can still be determined through watermark detection and tamper analysis.

[0037] Furthermore, compared to watermarking schemes used for tracing or detection in related technologies, the scheme of this disclosure does not aim for watermark recovery under all transformations, but rather adopts a semi-fragile approach: it remains robust to benign transformations such as compression and transcoding, ensuring stable watermark extraction under normal distribution scenarios. It exhibits vulnerability to malicious editing such as text overlay, image overlay, and cropping; that is, when such editing occurs, the neural network watermark verifier can detect and trigger failure. Exemplarily, this disclosure can be achieved through the following technical means: the neural network watermark verifier adopts a dual-head structure, adding a malicious editing detection head in addition to the watermark payload recovery head, specifically identifying pixel-level or region-level tampering behavior. During the training phase, data augmentation is divided into a benign transformation set (such as compression, blurring, and color adjustment) and a malicious editing set (such as text overlay, image patching, and local cropping), and differential supervision signals enable the model to produce discriminative responses to the two types of transformations. The payload recovery accuracy is optimized under benign transformations, and the tampering detection rate is optimized under malicious editing, while suppressing false positives for benign transformations. Furthermore, existing watermarking schemes employ a post-event detection model, with the detection results used for accountability rather than real-time protection. In this embodiment, a neural network watermarking verifier is integrated into the terminal playback link to achieve real-time verification frame-by-frame or sampled frame verification, and the verification results are directly linked to broadcast control actions. In addition, existing watermarking schemes often suffer from false alarms due to single-frame decoding anomalies and instantaneous frame drops during terminal deployment, making stable operation in real broadcast control environments difficult. Therefore, some embodiments of this disclosure also introduce a sliding window fault tolerance strategy. For example, within a time window of length N, blocking is only triggered if malicious editing is detected more than a threshold number M times, or if the number of watermark verification failures exceeds a threshold; occasional single-frame anomalies can be retried or downgraded to avoid playback interruptions due to occasional false alarms. This mechanism ensures security while achieving a controllable false alarm rate, making the scheme feasible for engineering deployment.

[0038] The solutions to the above problems will be explained in further detail below.

[0039] This disclosure provides a video playback control system that adopts a two-level architecture of server-side (or service center-side) and terminal-side to achieve a secure closed loop of server review, watermark embedding, content distribution, terminal verification, and broadcast control blocking.

[0040] Reference Figure 1 The diagram shown is a schematic representation of the overall architecture of a distributed video broadcast control content security system provided in this embodiment of the disclosure. Figure 1 As shown, the system adopts a two-tier architecture: a central side (service center, or server, which can be located in the cloud or a central data center) and a terminal side. The central side performs AI content review on the videos to be distributed. Once approved, an AI watermark is embedded, and the watermarked video is distributed to the terminals. The terminal side (broadcast controller / player) extracts video frames from the playback path, calls a neural network watermark verifier for watermark extraction and malicious editing detection, and uses a watermark verification and tampering judgment with a fault-tolerant mechanism (time window method, which will be explained in detail later). If the verification is successful and no tampering is detected, playback proceeds normally; otherwise, playback is prohibited and an evidence collection alarm is triggered. In the diagram, Y represents a yes result (approved, verification successful), and N represents a no result.

[0041] The video playback control method provided in this disclosure will be described in detail below from the perspectives of the broadcast control terminal side and the server side, respectively.

[0042] like Figure 2 As shown, this disclosure provides a video playback control method applied to a broadcast control terminal. The video playback control method includes the following steps: S201: Obtain the watermarked video file distributed locally. The watermarked video file is generated and distributed by the server after embedding a digital watermark into the approved video file using a neural network watermark embedder. The neural network watermark embedder is obtained by generating training video data based on benign transformations and malicious editing through dynamic data augmentation, and then co-training it with the neural network watermark verifier on the broadcast control terminal side by optimizing the joint loss function. The watermark embedded by the neural network watermark embedder is invisible to the naked eye, robust to preset benign transformations and can be successfully recovered by the neural network watermark verifier, and sensitive to malicious editing and can be detected by the neural network watermark verifier.

[0043] In this step, the broadcast control terminal first obtains the watermarked video file distributed to its local machine by the server (or service center). This watermarked video file is generated and distributed by the server after embedding a digital watermark into the approved original video file using a neural network watermark embedder. The watermark embedded by the neural network watermark embedder is invisible to the naked eye. The broadcast control terminal can obtain this watermarked video file through network reception, offline import, or local storage retrieval.

[0044] As an optional implementation, after acquiring the watermarked video file, the broadcast control terminal can first perform an integrity check on the file, such as through hash value comparison or digital signature verification, to ensure that the file has not been tampered with or damaged during transmission. If the integrity check fails, the broadcast control terminal can directly refuse playback and report the abnormal information. This integrity check step can effectively prevent the video file from being maliciously replaced or damaged during distribution and transmission, further enhancing security.

[0045] Through this step, the broadcast control terminal obtains the video content linked to the server's review conclusion, laying the foundation for subsequent playback verification. The server embeds the watermark only after the review is passed, making the watermark a carrier of the review conclusion and binding the review conclusion to the video content.

[0046] In this embodiment of the disclosure, the neural network watermark embedder and the neural network watermark verifier on the broadcast control terminal side are trained together by optimizing the joint loss function, so that the watermark generated by the neural network watermark embedder is robust to benign transformations including video encoding and decoding, image scaling, and resolution transformation, and sensitive to malicious editing including image overlay, partial content tampering, and cropping. At the same time, the neural network watermark verifier can accurately recover the watermark payload and detect malicious editing behavior.

[0047] S202: In the video playback processing link, a neural network watermark verifier is called to verify the playback frame and obtain the verification result; the neural network watermark verifier includes a watermark payload recovery unit and a malicious editing detection unit. The watermark payload recovery unit is used to recover the watermark payload from the playback frame, and the malicious editing detection unit is used to detect whether the playback frame has been maliciously edited; the verification result includes the watermark verification result and the malicious editing detection result.

[0048] In this step, the broadcast control terminal calls the neural network watermark verifier in real time in the video playback processing link to verify the video frames being played in order to determine whether the currently played content is the original version approved by the service center and whether malicious editing has occurred.

[0049] In this embodiment, the neural network watermark embedder and the neural network watermark verifier are a pair of jointly trained neural network models. The neural network watermark embedder is used to embed binary watermark messages into video frames to generate watermarked samples. Exemplarily, the neural network watermark embedder can adopt a dual-input UNET network structure, accepting a fixed-size image tensor input and a vector input of the watermark binary code. The watermark binary code is converted into a vector of the same dimension as the image features through a binary lookup table and copied and expanded along the spatial dimension. It is then spatially concatenated and added to the image features in the network to complete the embedding of the watermark information. The structure of the neural network watermark verifier includes a watermark payload recovery unit (or payload prediction head) and a malicious editing detection unit (or malicious editing detection head). The watermark payload recovery unit is used to recover the embedded watermark payload information from the input video frame, and the malicious editing detection unit is used to detect whether the input video frame has undergone malicious editing behavior, such as, but not limited to, text rendering overlay, image overlay, local cropping, trimming, patching, and other tampering operations.

[0050] There are several options for when to invoke the neural network watermark verifier in the video playback processing chain. As one possible implementation, the broadcast control terminal can invoke the neural network watermark verifier frame-by-frame for verification after video frame decoding and before rendering and display. This method enables the highest precision real-time monitoring, ensuring that each frame is verified before display, thus providing high security.

[0051] As an alternative implementation, considering the consumption of computing resources, the broadcast control terminal can also sample one frame every preset number of frames and call the neural network watermark verifier for verification. For example, one frame can be sampled for verification every 10 frames or every 30 frames. The specific sampling interval can be dynamically adjusted or pre-configured according to factors such as the computing power of the terminal device, video resolution, and security level requirements. This sampling verification method significantly reduces the terminal's computing overhead while ensuring basic security, which is beneficial for deployment on terminal devices with limited computing power.

[0052] In optional implementations, to reduce the deployment overhead of the neural network watermarking validator on computationally limited terminals, lightweight model techniques and hierarchical verification strategies can be employed. For example, the neural network watermarking validator can be compressed using one or more of the following lightweight techniques: model distillation, using a lightweight student network to mimic the output of the teacher network, reducing the number of model parameters and computational load; quantization, converting model weights from floating-point to low-bit integers (such as INT8), reducing storage and computational overhead; structural pruning, removing redundant channels or layers from the model; and an early stopping mechanism, exiting the inference process early, ceasing further computation when intermediate features are sufficient to support the decision. Another example is deploying a lightweight neural network watermarking validator on the terminal side as a primary screening model to quickly verify all playback frames. When the confidence level of the primary screening model's output is lower than a preset threshold or an anomaly is suspected, the frame is then sent to a more accurate but computationally more computationally expensive secondary verification model for review, and a final broadcast control decision is made based on the review results. This hierarchical strategy can significantly reduce the average computational load on the terminal side while ensuring security.

[0053] In some embodiments, the watermark payload recovery unit (payload prediction head) and the malicious editing detection unit (malicious editing detection head) can share the underlying feature extraction network. In this case, the process of calling the neural network watermark verifier to verify the playback frame may include: inputting the playback frame into the neural network watermark verifier to obtain the watermark payload output by the watermark payload recovery unit, and the tampering heatmap or tampering score output by the malicious editing detection unit. If the tampering score is greater than a preset tampering score threshold, or the maximum value of the tampering heatmap is greater than a preset maximum value threshold, or the average value of the tampering heatmap is greater than a preset average value threshold, then malicious editing is determined to be detected; if the confidence level, bit error rate, or p-value of the watermark payload does not meet the corresponding threshold, then watermark verification is determined to have failed.

[0054] For example, the malicious editing detection is first performed: if the score is tampered with... Greater than the preset tampering score threshold , or the maximum value of the tampered heatmap Greater than the preset maximum threshold , or the average value of the tampered heatmap Greater than the preset average threshold If malicious editing is detected, the watermark verification will fail.

[0055] If none of the above malicious editing detections are triggered, the process proceeds to the watermark payload determination stage. For example, the watermark payload determination can be performed through at least one of the following methods: (1) Determination based on the confidence level of the watermark payload: the soft bit vector output by the watermark payload recovery unit This represents the probability that each bit is 1. The binary watermark payload is obtained by thresholding the soft bit vector. Then, the average confidence level of each bit can be calculated. If the average confidence level is lower than the preset confidence threshold, it indicates that the reliability of the watermark extraction result is insufficient, and the watermark verification is deemed to have failed.

[0056] (2) Judgment based on bit error rate: Bit error rate (BER) refers to the ratio of errors between the extracted watermark bits and the original watermark bits. If the terminal knows the original watermark content in advance (e.g., the watermark is a fixed identifier or can be inferred from the content identifier), it can calculate the bit error rate between the extracted watermark and the original watermark. If the bit error rate exceeds a preset threshold (e.g., more than 10%), it indicates that the watermark content does not match the expectation, and the watermark verification is judged to have failed.

[0057] (3) Judgment based on p-value: The p-value is used to measure the statistical significance of watermark extraction results, representing the probability of observing the current extraction result under the assumption of no watermark. When the p-value is sufficiently small, it indicates that the evidence of watermark existence is sufficient; when the p-value is large, it indicates that the extraction result may be caused by random noise. For example, the p-value can be set to satisfy... This serves as a condition for watermark verification to pass; if this condition is not met, the watermark verification is deemed to have failed.

[0058] If the watermark payload confidence level meets the preset threshold, the bit error rate does not exceed the preset threshold, and the p-value meets at least one of the preset thresholds (depending on the actual judgment method used), then the watermark verification is deemed to have passed; otherwise, the watermark verification is deemed to have failed.

[0059] It should be noted that the above three determination methods can be selected or combined according to the actual application scenario. For example, in scenarios where it is only necessary to determine whether the watermark exists, confidence level determination or p-value determination can be used; in scenarios where it is necessary to verify the correctness of the watermark content, bit error rate determination can be used.

[0060] Here, the output of the watermark payload recovery unit can be a soft bit vector or directly recovered binary watermark information, while the output of the malicious editing detection unit can be a pixel-level tampering heatmap. Each pixel value represents the probability that the location has been tampered with, suitable for scenarios requiring precise location of the tampered area. It can also output a global tampering score. This represents the overall probability of malicious editing of the entire image frame, suitable for terminal devices with limited computing power. It can also output patch-level tampering heatmaps, dividing the image into multiple patches, with each patch outputting a tampering probability, reducing computational load compared to pixel-level output. When using a global tampering score, the malicious editing detection unit can employ an attention mechanism to aggregate feature maps, such as mapping the feature maps to scalar scores through spatial attention-weighted averaging or self-attention pooling layers, to improve the accuracy of tampering detection and adapt to edge computing power.

[0061] The above threshold parameters ( , The confidence threshold, bit error rate threshold, and p-value threshold can be flexibly configured according to the security level requirements of the actual application scenario. For example, in scenarios with high security requirements, such as public displays, stricter thresholds can be set to ensure that any slight anomaly can be detected and trigger a blocking mechanism. In ordinary video playback scenarios, the thresholds can be appropriately relaxed to avoid false alarms caused by occasional noise that could affect the user experience.

[0062] In some embodiments, to achieve a semi-fragile characteristic of being robust to benign transformations and sensitive to malicious editing, the neural network watermarking embedder and the neural network watermarking verifier employ a joint training strategy. During joint training, the training data uses a separate augmentation approach, dividing it into a benign transformation set and a malicious editing set. Specifically, the neural network watermarking verifier and the neural network watermarking embedder can be jointly trained as follows: Training video data is acquired, and the training augmentation is divided into a benign transformation set and a malicious editing set; the benign transformation set includes video encoding, video transcoding, image quality degradation, color or brightness adjustment, and minor geometric transformations; the malicious editing set includes text rendering overlay, image overlay, local patching, and cropping or trimming; transformations from the benign transformation set are applied to the training video data to obtain benign transformation samples; transformations from the malicious editing set are applied to the training video data, and pixel-level mask labels are generated to obtain malicious editing samples and their corresponding pixel-level mask labels; based on the benign transformation samples, the malicious editing samples, and their corresponding pixel-level mask labels, the neural network watermarking verifier and the neural network watermarking embedder are jointly trained.

[0063] Reference Figure 3 The diagram shown is a schematic of a dual-head decoding network structure with malicious editing detection provided in an embodiment of this disclosure. Figure 3As shown, the network includes a feature encoder, residual blocks, average pooling layers, a multilayer perceptron (MLP), and convolutional layers. The input image tensor (3×256×256) is processed by the feature encoder to extract features. One path is processed by the residual block, average pooling, and MLP to output a watermark vector (256 bits). The other path is processed by the residual block and a 1×1 convolution to output a malicious editing probability map (1×16×16), achieving dual-head output for watermark payload recovery and malicious editing detection.

[0064] For example, training video data is obtained, and training enhancements are divided into a set of benign transformations and a set of malicious edits.

[0065] Among them, the benign transformation set This can include the following transformation types: Video encoding / transcoding: including encoding standards such as H.264, H.265, and AV1, randomly setting parameters such as bitrate, CRF (Constant Quality Factor), or QP (Quantization Parameter) to simulate the compression processing that video may undergo during distribution and transmission. Image quality degradation: including Gaussian blur, median filtering, slight noise addition, slight sharpening, etc., simulating the quality changes that may occur during image transmission or processing. Color or brightness adjustment: including slight perturbations in brightness, contrast, saturation, and hue to simulate color differences under different display devices or playback environments. Slight geometric transformations: including small-angle rotation, slight scaling, slight perspective perturbations, etc., simulating slight deformations that may occur during video playback or acquisition.

[0066] The goal of the aforementioned benign transformation is to simulate normal processing that may occur during real distribution and transmission without changing the semantics of the content. The trained model can stably recover the watermark payload under these transformations, while keeping the output of the tamper detection unit close to all zero, i.e., not falsely reporting it as malicious editing.

[0067] Malicious Edit Collection The following transformation types are available: TextOverlay: Randomly overlays text content onto video frames to simulate badges, subtitles, bullet comments, watermark-style text, etc. Text parameters can be randomly set, including but not limited to font, font size, color, stroke, shadow, transparency, rotation angle, and position. PatchOverlay: Pastes tiles, logos, or overlays onto local areas of video frames to simulate advertising patches, badge overlays, or content occlusion. Inpaint / BlurPatch: Blurs, mosaics, or repairs images on local areas of video frames to simulate content alteration or occlusion in specific areas. Crop / Cut / Paste: Crops video frames and scales them back to their original resolution, or cuts and pastes local areas to simulate deletion, replacement, or splicing of content.

[0068] The goal of the aforementioned malicious editing transformation is to simulate real malicious tampering behavior. The trained model can accurately identify these tamperings and output the corresponding tampering signals.

[0069] For malicious edits that can be synthesized (such as text rendering overlays, image overlays), output pixel-level mask labels while generating the transformation. The edited area is labeled with 1, and the remaining areas are labeled with 0, serving as the supervisory label for the tampering detection unit. It is important to note that text contained in the original video frame itself (as part of the pixel content) is not labeled as tampered; only newly added, covered, or modified areas are labeled as tampered, thus avoiding misjudging native text content in the video as malicious editing.

[0070] After applying the above transformations to the training video data, the following samples are obtained: Applying transformations from the benign transformation set to the training video data yields benign transformation samples. This is used to train the model's robustness under benign processing. Transformations from a malicious editing set are applied to the training video data, and corresponding pixel-level mask labels are generated to obtain malicious editing samples. and the corresponding pixel-level mask labels This is used to train the model's sensitivity to malicious editing.

[0071] Through the aforementioned training data augmentation strategy, the neural network watermarking verifier can learn to distinguish between benign transformations and malicious edits, thereby achieving a semi-fragile property that is robust to benign transformations and sensitive to malicious edits.

[0072] In some embodiments, during joint training, the neural network watermark embedder is used to embed binary watermark messages into video frames to generate watermarked samples; the neural network watermark verifier is used to recover the watermark payload from the watermarked samples and detect malicious editing. The neural network watermark verifier can be trained using a multi-task loss function for joint optimization, for example, it may include: Based on the benign transformation samples, the watermark payload recovery loss, the imperceptible constraint loss, and the false alarm suppression loss of the neural network watermark verifier and the neural network watermark embedder are calculated. The false alarm suppression loss is used to constrain the tampering heatmap output by the malicious editing detection unit under benign transformation to be close to all zeros. Based on the malicious editing samples and the corresponding pixel-level mask labels, the tampering detection loss of the malicious editing detection unit of the neural network watermark verifier is calculated. The watermark payload recovery loss, the imperceptible constraint loss, the false alarm suppression loss, and the tampering detection loss are weighted and summed to obtain a joint loss function. The neural network watermark verifier and the neural network watermark embedder are jointly trained by optimizing the joint loss function.

[0073] For example, firstly, based on the benign transformation sample, the following loss is calculated: (1) Watermark load recovery loss : Samples after benign transformation The goal is to minimize the binary cross-entropy loss between the predicted watermark payload and the original watermark payload. This loss function can be expressed as: ; in, This is the predicted watermark payload (soft bit vector) output by the watermark payload recovery unit for the benignly transformed sample. The original watermark payload (binary message) is represented by BCE, which denotes the binary cross-entropy loss. This loss term ensures that the model can accurately recover the watermark information under benign transformations such as compression and transcoding, achieving robustness to benign processing.

[0074] (2) Loss due to imperceptible constraints : Watermarked frames output by the embedder With the original frame To minimize the mean squared error loss and ensure that watermark embedding does not affect the visual quality of the video, the loss function can be expressed as: ; in, This represents the square of the L2 norm, which is the sum of squares of the pixel-by-pixel differences. This loss term makes the watermark visually imperceptible, ensuring that the user experience is not affected. In an alternative implementation, perceptual loss can be used instead of mean squared error loss to better align with the characteristics of human vision.

[0075] (3) False alarm suppression loss : For benign transformation samples To constrain the malicious editing detection unit, the output heatmap of the tampering is kept close to all zeros, thus preventing the model from misclassifying benign transformations as malicious editing. This loss function can be expressed as: ; in, This is a tampering heatmap output by the malicious editing detection unit for benignly transformed samples, where 0 represents an all-zero mask. This loss term is key to achieving the semi-fragile property; it forces the model not to output tampering signals under benign transformations, thereby effectively reducing the false alarm rate and ensuring that normal video compression, transcoding, and other processing will not trigger security blocking.

[0076] Secondly, based on the maliciously edited sample and the corresponding pixel-level mask label, the tamper detection loss is calculated. : ; in, This is a predicted tampering heatmap output by the malicious editing detection unit for maliciously edited samples. The loss term provides pixel-level mask labels (edited areas are labeled with 1, and the rest with 0). This loss term enables the model to accurately identify maliciously edited areas, achieving pixel-level tamper location capabilities.

[0077] Then, the weighted sum of the above losses yields the joint loss function:

[0078] in, These are the weighting coefficients for each loss term, used to balance the importance of different tasks in joint training. The values ​​of these weighting coefficients can be adjusted based on the actual training results. For example, they can be set to... , , , This makes watermark payload recovery and tamper detection the main optimization goals, while also taking into account visual quality and false alarm suppression.

[0079] During training, an adaptive balancing strategy based on gradient norm can be used to dynamically adjust the weight coefficients to stabilize the training process. Specifically, the gradient norm of each loss term can be monitored. When the gradient norm of a certain loss term is too large, its weight is appropriately reduced, and vice versa, so that the various losses remain balanced during training and prevent any one task from dominating the training direction.

[0080] Finally, by optimizing the joint loss function, the neural network watermark validator is trained, enabling the model to simultaneously possess the following capabilities: For benignly transformed samples: accurately recover the watermark payload (through... Optimization), keeping the watermark imperceptible (through (Optimized), and does not output tampering signals (through) (Optimization). For maliciously edited samples: It can accurately identify the tampered areas and output a tampering heatmap (via...). optimization).

[0081] Through the above multi-task joint optimization training, the neural network watermarking verifier achieves a semi-fragile property that is robust to benign transformations and sensitive to malicious editing, providing a reliable model foundation for real-time verification on the terminal side.

[0082] In one optional implementation, to further optimize training performance and achieve semi-fragile characteristics, the joint training of the neural network watermark embedder and the neural network watermark verifier can employ a three-stage training strategy, including: The training process consists of three phases: Phase 1: Using image data as training samples, the neural network watermark embedder and the neural network watermark validator are jointly trained, optimizing the watermark payload recovery loss and imperceptible constraint loss in the joint loss function. Phase 2: Using both image and video data as training samples, a benign transformation related to video compression and a time-based watermark propagation strategy are employed for training. Maliciously edited samples and their pixel-level mask labels are also used for training, optimizing the tamper detection loss and false alarm suppression loss in the joint loss function. Phase 3: The network parameters of the neural network watermark embedder are frozen, and the neural network watermark validator is fine-tuned, optimizing the tamper detection loss and false alarm suppression loss in the joint loss function.

[0083] In the three-stage training process described above, the first stage involves image pre-training. This stage utilizes a large amount of image data to jointly pre-train the neural network watermark embedder and validator. This stage primarily optimizes the watermark payload recovery loss and the imperceptibility constraint loss, enabling the embedder to generate visually high-quality watermarks, while the validator can accurately recover the watermark payload from the image, laying the foundation for subsequent training. The second stage involves mixed image / video training. In this stage, video data is introduced, and temporally related benign transformations such as video encoding and transcoding, as well as a time-based watermark propagation strategy, are incorporated for training. Simultaneously, maliciously edited samples (such as text overlays, image overlays, etc.) and their mask labels are added to the training data. This stage optimizes the joint loss function, including tamper detection loss and false positive suppression loss, enabling the validator to learn to distinguish between benign transformations and malicious editing, initially acquiring semi-fragile detection capabilities. In the third stage, the embedder is frozen, and the validator is fine-tuned. The network parameters of the neural network watermark embedder trained in the first stage are frozen, and only the neural network watermark validator is fine-tuned. This stage uses more challenging maliciously edited samples (such as smaller text and patches more similar to the background texture) for training, and focuses on optimizing the tamper detection loss and false positive suppression loss. This stage aims to further improve the validator's sensitivity to malicious editing, while strictly controlling the false positive rate for benign transformations, and without changing the visual appearance of the embedded watermark.

[0084] S203: If the verification result indicates that the verification is successful, continue video playback; if the verification result indicates that the verification is unsuccessful, execute the broadcast control blocking action and report the failure information.

[0085] In this step, the broadcast control terminal determines its subsequent actions based on the verification result obtained in S202. If the verification result indicates that the verification is successful, such as the watermark verification is successful and no malicious editing is detected, the broadcast control terminal continues to play the video content normally; if the verification result indicates that the verification is unsuccessful, the broadcast control terminal immediately executes a broadcast control blocking action and simultaneously reports the failure information to the server for evidence collection and auditing.

[0086] In some embodiments, a sliding window fault-tolerance strategy can be used to determine whether to trigger blocking. For example, the verification results of multiple playback frames within a preset time window (e.g., 10 to 100 frames selected according to actual needs) are obtained, and the number of frames determined to be maliciously edited or the number of frames that failed watermark verification within the preset time window are counted; if the counted result exceeds a preset threshold, it is determined that the verification result indicates that the verification failed, and then the broadcast control blocking action is triggered and the failure information is reported. Here, the preset threshold can be a frame number threshold or a ratio threshold (e.g., 0.3 to 0.7).

[0087] For example, a time window (sliding window) of length N frames can be set to record and statistically analyze the verification results of each frame within the window. Let each frame yield a tampering score. (The maximum value of the heatmap can be modified) or average (Acquired by aggregation), confidence level or p-value index obtained by watermark load detection. Within a sliding window of length N, if the following conditions are met... This means that the score was tampered with beyond the threshold. If the frame rate reaches M frames, then malicious editing will fail; otherwise, if If the watermark payload fails to trigger, the watermark payload will fail to load; otherwise, the test will pass. For indicator functions, indicating if If the count is 1, then the count is 1; otherwise, the count is 0. For indicator functions, indicating if If it does, the count is 1; otherwise, the count is 0.

[0088] Among them, N, M, These are configurable hyperparameters that can be adjusted according to the security level and false alarm tolerance of the actual application scenario. For example: N=30, M=3. =3: This indicates that within a 30-frame time window (approximately 1 second, assuming 30fps), if the number of frames detected as maliciously edited reaches 3, or the number of frames failing watermark verification reaches 3, then broadcast control blocking will be triggered. In scenarios with high security requirements, smaller N and M values ​​can be set to achieve a faster response; in scenarios with relatively relaxed security requirements, larger N and M values ​​can be set to reduce the false alarm rate.

[0089] If a single frame fails occasionally (e.g., watermark extraction fails due to a momentary decoding anomaly), since the cumulative threshold condition within the sliding window is not met, a retry or downgrade process can be selected, without triggering playback control blocking, thereby avoiding playback interruption due to momentary anomalies.

[0090] This sliding window fault-tolerance mechanism ensures security while achieving a controllable false alarm rate, making this solution feasible for engineering deployment. Specifically: for continuous malicious editing (such as superimposing text or badges across multiple consecutive frames), the number of maliciously edited frames within the sliding window will quickly accumulate and exceed the threshold M, thereby triggering broadcast control blocking to ensure that abnormal content is not broadcast. For occasional anomalies (such as single-frame decoding errors, momentary screen tearing caused by network jitter, etc.), since the accumulation threshold condition is not met, broadcast control blocking will not be triggered, ensuring playback continuity and user experience. By adjusting N, M, ... These parameters can strike a balance between security and user experience, adapting to the needs of different application scenarios.

[0091] In one alternative implementation, the sliding window can employ a sliding window mechanism, where the window slides forward one frame after each frame is processed, updating the statistical results within the window in real time. This approach enables continuous monitoring and timely detection of persistent anomalies.

[0092] In another alternative implementation, a segmented window mechanism can be used, where each N frames is treated as an independent detection window, and the windows do not overlap. This approach is simpler to implement, has lower computational overhead, and is suitable for scenarios where real-time requirements are not extremely high.

[0093] The aforementioned sliding window fault-tolerance mechanism is one of the key design features for achieving engineering deployment in this application. By combining the single-frame verification result with the cumulative statistics within the time window, it ensures that malicious editing can be detected and blocked in a timely manner, while effectively avoiding false alarms caused by occasional anomalies, enabling the system to operate stably in a real and complex broadcast control environment.

[0094] In some embodiments, executing the broadcast control blocking action and reporting failure information includes: executing the broadcast control blocking action, which includes at least one of pausing playback, displaying a black screen, cutting off the video output signal, or switching to backup video content. Simultaneously, the content identifier, timestamp, failure type, frame summary information, and terminal device identifier of the playback frame are extracted, failure information is generated, and reported to the server; wherein, the failure type includes watermark verification failure and / or detection of malicious editing. Through the above-mentioned reported information, the server can promptly obtain abnormal situations and perform auditing and tracing, facilitating subsequent analysis and processing.

[0095] For example, the execution of the broadcast control blocking action and the reporting of failure information includes executing the broadcast control blocking action and simultaneously extracting and reporting the failure information.

[0096] Specifically, for executing broadcast control blocking actions: when the verification result indicates watermark verification failure or malicious editing is detected, the broadcast control terminal immediately executes a broadcast control blocking action to prevent abnormal content from being broadcast. The broadcast control blocking action includes, but is not limited to, at least one of the following methods: 1) Pause playback: Immediately pause video playback, remaining on the current screen or displaying a pause prompt, awaiting subsequent processing instructions. 2) Black screen display: Immediately switch the display screen to black, completely blocking the visual output of abnormal content, while simultaneously displaying security prompt information. 3) Cut off video output signal: Cut off the video output signal at the hardware or system level, making the display device have no signal input, achieving physical-level playback blocking. 4) Switch to backup video content: Immediately switch to a pre-configured backup stream, padding content, or default video, ensuring uninterrupted broadcast while preventing the abnormal content from being broadcast. The above broadcast control blocking actions can be flexibly configured according to actual application scenarios. For example, in a public display screen scenario, it can be set to immediately black out the screen and display a prompt message, such as indicating that the content security verification failed, when an anomaly is detected, both blocking the abnormal content and explaining the situation to the audience. In live streaming scenarios, the system can be configured to immediately switch to a backup stream or filler content upon detecting an anomaly, ensuring uninterrupted broadcast while preventing the abnormal content from being played. In unattended outdoor advertising screen scenarios, the system can be configured to immediately cut off the video output signal upon detecting an anomaly, simultaneously sending an alarm to the management platform and awaiting remote recovery. In government and enterprise video publishing scenarios, the system can be configured to pause playback, record anomaly information, and enter a manual confirmation mode, allowing administrators to determine subsequent handling methods.

[0097] For synchronous extraction and reporting of failure information: While executing the broadcast control blocking action, the broadcast control terminal synchronously extracts various information related to this failure, generates failure information, and reports it to the server for evidence collection, auditing, and subsequent analysis. The failure information includes, but is not limited to, the following: 1) Content identifier: A unique identifier for the currently playing video (such as content ID, video number, etc.), used to locate which specific video file experienced the anomaly. 2) Timestamp: The precise time of the failure (such as the frame timestamp or system time at the time of playback), used to trace the specific time point of the anomaly. 3) Failure type: Identifies the specific reason for this failure, including watermark verification failure and / or detection of malicious editing. Further, failure types can be subdivided into: Watermark verification failure: including subtypes such as watermark payload recovery failure, insufficient watermark payload confidence, p-value index not meeting the threshold, and bit error rate exceeding the threshold. Detection of malicious editing: including the identification results of specific tampering types such as text rendering overlay, image overlay, local repair, and cropping / cutting. 4) Frame Summary Information: Relevant information about the failed playback frame, including: thumbnails or feature vectors of keyframes for manual review or automated analysis; visualization of the tampering heatmap, marking the location of areas detected as maliciously edited; and frame sequence numbers for locating abnormal locations in the video. 5) Terminal Device Identifier: A unique identifier for the broadcast control terminal (e.g., device ID, MAC address, IP address), used to pinpoint which terminal experienced the anomaly, facilitating terminal-level statistical analysis and troubleshooting.

[0098] Furthermore, after the broadcast control terminal reports the aforementioned failure information to the server, the server can, for example, perform the following operations: 1) Real-time alarm: For scenarios with high security requirements, the server can immediately send alarm notifications to administrators (such as SMS, email, or APP push notifications) for timely handling. 2) Audit and traceability: The failure information is stored in the audit log, recording detailed information such as the time, location, content, and cause of the anomaly, facilitating post-event traceability and responsibility identification. 3) Statistical analysis: The reported failure information is categorized and statistically analyzed to determine the distribution of anomalies across different terminals, content, and time periods, providing data support for system optimization and security policy adjustments. 4) Automatic recovery: For anomalies that can be automatically recovered (such as watermark extraction failure caused by network jitter), the server can issue recovery instructions to the terminal or redistribute the correct video content. Through this reporting mechanism, the server can promptly obtain anomalies and perform audit and traceability, facilitating subsequent analysis and processing.

[0099] In some optional implementations, the broadcast control blocking action, in addition to immediately pausing playback, displaying a black screen, cutting off the video output signal, or switching to backup video content, can also include a variety of other optional handling strategies. For example, a short pause and retry can be performed: when the verification result is an occasional failure of a single frame, the broadcast control terminal pauses playback and waits briefly (e.g., 100-300 milliseconds) to retry verification on subsequent frames. If the verification passes continuously during the retry process, playback is resumed; if the cumulative failures within the retry window reach a threshold, permanent blocking is triggered. This strategy can effectively deal with occasional verification failures caused by instantaneous decoding anomalies or network jitter, avoiding false blocking. Another example is entering a manual confirmation mode: when verification fails and the security level requirement is high, the broadcast control terminal can pause playback and enter a manual confirmation mode, reporting the failure information to the management platform. Administrators can then remotely confirm and decide whether to resume playback, switch content, or continue the blocking. This is suitable for scenarios requiring manual review, such as government and enterprise releases and public screen broadcast control. For example, differentiated handling can be implemented for internal and external components: the broadcast control terminal can block external output signals (such as blacking out the screen or cutting off the output), while continuing to play and record internally for evidence collection. The complete recording of the abnormal content can be saved locally or uploaded to the server for subsequent auditing, tracing and responsibility determination, thus meeting the dual needs of security blocking and evidence preservation.

[0100] like Figure 4 As shown in the figure, this disclosure also provides a video playback control method, applied to a server or service center, which includes the following steps: S401: Obtain the video file to be published, conduct content review on the video file to be published, and output the review conclusion.

[0101] In this step, the server obtains the video file to be published and calls the AI ​​content review module to conduct a comprehensive review of the video file to determine whether its content complies with broadcast regulations. The review conclusion may include whether the review is passed or failed, and may also include additional information such as policy version number, review time, and content identifier. Only videos that pass the review proceed to the subsequent watermark embedding process; videos that fail the review are prohibited from distribution or enter the manual review process.

[0102] In some embodiments, the content review of the video file to be published may include: extracting keyframes from the video file to be published, calling a visual content review model to perform image-level review, and obtaining an image review result; extracting the audio stream from the video file to be published and converting it into text, calling a text content review model to perform text-level review, and obtaining a text review result; based on the image review result and the text review result, comprehensively determining the compliance of the video file to be published, and outputting a review conclusion. For example, it can be set that the review is only considered passed when both the image review and the text review are passed; if either review fails, the review is considered failed. This multimodal review method can comprehensively cover the compliance detection of video content, improving the accuracy and completeness of the review.

[0103] S402: If the review conclusion indicates that the review is passed, then the neural network watermark embedder is invoked to embed a neural network digital watermark in the frames of the video file to be published, thereby generating a watermarked video file.

[0104] In this step, if the review conclusion output by S401 indicates that the review has passed, the server embeds a neural network digital watermark into the video frames of the video file to be published, generating a watermarked video file. The watermark embedding process binds the review conclusion to the video content, allowing subsequent terminal playback to verify the authenticity and completeness of the content through watermark verification. If the review conclusion indicates that the review has failed, the distribution of the video file to be published is prohibited, or it is sent to a manual review process.

[0105] In some embodiments, embedding a neural network digital watermark in the frames of the video file to be published includes: in response to the review conclusion indicating that the review has been passed, extracting the content identifier, policy version number, and review time from the review conclusion and encoding them into a binary message; embedding the binary message as watermark information into the frames of the video file to be published, so that the watermarked video file carries the unique identifier of the review conclusion.

[0106] In this way, the terminal can not only determine whether the watermark exists during verification, but also verify whether the content identifier carried in the watermark matches the expectation, further enhancing the reliability of the verification.

[0107] In practical implementation, in response to the review conclusion indicating that the review has passed, key information from the review conclusion is extracted. For example, when the server reviews a video file, if the review conclusion is "approved," the server extracts key information from the review conclusion that can be used to identify this review. This key information includes, but is not limited to: Content Identifier (Content ID): a number or identifier used to uniquely identify the video file, facilitating subsequent traceability and management. Policy Version Number: identifies the version information of the content review policy used in this review. When the review policy is updated, the policy version number can be used to distinguish content approved at different times. Review Time: records the timestamp of this review approval, used for subsequent timeliness judgment (for example, it can be set that if the video is not played after a certain period of time after approval, it needs to be re-reviewed). The above key information can be further expanded. In optional implementations, additional information such as reviewer identifier, review organization identifier, and content source identifier can also be included to meet the traceability needs of different application scenarios.

[0108] Next, the extracted key information is encoded into a binary message. For example, the server converts the aforementioned key information into a binary message according to a preset encoding rule, serving as the watermark payload to be embedded. Encoding methods can include fixed-length encoding, variable-length encoding, or hybrid encoding. For instance, the content identifier uses 32-bit binary encoding; the policy version number uses 8-bit binary encoding; and the review time uses 32-bit binary encoding (such as a Unix timestamp). The total length of the binary message can be determined according to actual needs, such as 64 bits, 128 bits, or 256 bits. The length of the watermark payload determines the amount of information it can carry and also affects the robustness and imperceptibility of the watermark; a trade-off can be made based on the application scenario.

[0109] Then, the binary message is embedded as watermark information into the frames of the video file. For example, the server calls a neural network watermark embedder to embed the aforementioned binary message into the video frames of the video file. The watermark embedding method can employ the aforementioned encoder-decoder-based architecture, embedding the binary message in the pixel domain. Specifically, the watermark embedding process may include: scaling the video frame to a preset resolution (e.g., 256×256) to obtain a scaled frame; calling the neural network watermark embedder to embed the binary message in the pixel domain of the scaled frame, generating watermark distortion; upsampling the watermark distortion back to the original resolution and superimposing it onto the original frame with a preset intensity coefficient to obtain a watermarked video file. Through the above embedding process, the generated watermarked video file carries a unique identifier of the review conclusion, including key information such as content identifier, policy version number, and review time.

[0110] Next, content identifier verification is performed during terminal verification. For example, when the broadcast control terminal obtains a watermarked video file and performs playback verification, the neural network watermark verifier recovers the watermark payload (i.e., the aforementioned binary message) from the playback frame. The terminal decodes the recovered binary message to obtain information such as the content identifier, policy version number, and review time. The terminal compares the decoded content identifier with the expected value (e.g., the content identifier obtained from the playlist sent by the server): if they match, it indicates that the currently playing content is indeed the original version that has passed review, and the verification passes. If they do not match, it indicates that the watermark content does not match the expectation, and content replacement or tampering may have occurred; the watermark verification is deemed to have failed, triggering broadcast control blocking.

[0111] In this way, the terminal can not only determine the presence of a watermark during verification, but also verify whether the content identifier carried in the watermark matches expectations, further enhancing the reliability of the verification. Furthermore, the policy version number can be used to determine whether the review policy is outdated, and the review time can be used to determine whether the content has exceeded its validity period, providing a foundation for more granular security policies.

[0112] In an optional implementation, to enhance the detection capability of temporal tampering behaviors such as frame insertion, frame deletion, and splicing, the watermark payload can be designed as a structured field, and a temporal consistency verification mechanism can be introduced. For example, in addition to carrying content identifiers, policy version numbers, and review times, the watermark payload can also carry frame sequence numbers or segment sequence numbers. For instance, segment sequence numbers are assigned to video sequences at fixed intervals (such as per second or per group of frames (GOPs), and intra-frame sequence numbers are assigned to each frame, which are then encoded and embedded with the watermark. During verification, the broadcast control terminal recovers the watermark payload from continuously played frames, decodes it to obtain the frame sequence number or segment sequence number, and verifies its continuity: if the sequence numbers of adjacent frames are not consecutive, repeat, or show abnormal jumps, it is determined that frame insertion, frame deletion, or splicing tampering may have occurred, triggering a watermark verification failure.

[0113] Furthermore, a chain-based digest enhancement mechanism can be employed. The watermark payload of the current frame is hash-bound with the watermark payload of the previous frame, forming a chain-based verification structure. Specifically, the... Frame-embedded watermark payload May contain the first The hash value or digest information of the frame watermark payload. During terminal verification, in addition to restoring the watermark payload of the current frame, it is also necessary to verify whether its hash association with the previous frame is valid. If the association is broken, it is determined that frame-level tampering has occurred. This mechanism can effectively enhance the sensitivity to attacks on video sequence integrity.

[0114] In some embodiments, embedding a neural network digital watermark in the frames of the video file to be published includes: scaling the frames of the video file to be published to a preset resolution to obtain scaled frames; calling a neural network watermark embedder to embed binary messages in the pixel domain of the scaled frames to generate watermark distortion; upsampling the watermark distortion back to the original resolution and superimposing it onto the original frame with a preset intensity coefficient to obtain a watermarked video file.

[0115] This embedding method unifies the watermark embedding operation at a fixed resolution, greatly reducing the computational complexity of embedding, while ensuring the imperceptibility of the watermark through upsampling and overlay.

[0116] Reference Figure 5 The diagram shown is a schematic representation of the neural network watermark embedding and extraction process provided in an embodiment of this disclosure. Figure 5 As shown, the watermark embedding process includes: scaling the original frame to a fixed size, inputting it into the watermark embedding network to generate watermark distortion, upsampling the watermark distortion to the original image size and then superimposing it onto the original frame to obtain a watermarked frame; the watermark extraction process includes: scaling the watermarked frame to a fixed size, inputting it into the watermark extraction network, outputting a predicted watermark vector and a malicious editing score, thresholding the predicted watermark vector to obtain the watermark code, and using the malicious editing score to determine whether malicious editing exists.

[0117] In specific implementation, a neural network digital watermark is embedded in the frames of the video file to be published, using an encoder-decoder framework to embed a binary message payload in the pixel domain. For example, firstly, the frames of the video file to be published are scaled to a preset resolution to obtain scaled frames. The server acquires each frame of the video file to be published. Since the original video frame resolution may be high (e.g., 1920×1080, 3840×2160, etc.), directly embedding the watermark at high resolution would incur significant computational overhead. To reduce computational complexity, the server first scales the original frames to a fixed preset resolution to obtain scaled frames. This preset resolution can be set according to the design of the neural network watermark embedder, for example, it can be set to a smaller size such as 256×256 or 224×224. By scaling the high-resolution frames to a fixed low resolution, subsequent watermark embedding operations are performed at a uniform low resolution, greatly reducing the computational load of the neural network and improving embedding efficiency. Simultaneously, the fixed-resolution input simplifies the design of the neural network model, allowing the model to focus on learning the mapping relationship of watermark embedding.

[0118] Then, a neural network watermark embedder is invoked to embed the binary message in the pixel domain of the scaled frame, generating watermark distortion. For example, the server inputs the scaled frame into a pre-trained neural network watermark embedder. This neural network watermark embedder is a deep learning model that takes the scaled frame and the watermark payload (binary message) as input and outputs the watermark distortion. The watermark distortion is a residual image of the same size as the scaled frame, representing the amount of modification needed in the pixel domain to embed the watermark. Specifically, the neural network watermark embedder can learn to map the binary message to a watermark distortion that adapts to the image content through an encoder-decoder structure. The watermark distortion has the following characteristics: in textured areas, the watermark distortion can be large but not easily noticeable; in flat areas, the watermark distortion is automatically suppressed to ensure visual quality; the watermark distortion is related to the image content, making the watermark highly imperceptible and robust.

[0119] Finally, the watermark distortion is upsampled back to the original resolution and superimposed on the original frame with a preset intensity coefficient to obtain the watermarked video file. For example, the server upsamples the generated watermark distortion (size at a preset resolution, such as 256×256) back to the resolution of the original frame (such as 1920×1080). Upsampling can use traditional image interpolation methods such as bilinear interpolation or bicubic interpolation, or it can use a learned upsampling module to ensure the smoothness and continuity of the watermark distortion. Then, the server superimposes the upsampled watermark distortion onto the original frame with a preset intensity coefficient α to obtain the final watermarked frame. This process can be represented as: ;in, For the original frame, To compensate for watermark distortion generated at scaled resolutions, For upsampling operation, The strength coefficient, This is the final output watermarked frame. The intensity coefficient α controls the balance between the visibility and robustness of the watermark: the larger α is, the stronger the watermark robustness, but it may affect visual quality; the smaller α is, the better the watermark's imperceptibility, but robustness may decrease. The value of α can be set according to the actual application scenario. For example, it can be set to a value between 0.1 and 0.3 to obtain sufficient robustness while ensuring imperceptibility.

[0120] Through the above process, the server performs watermark embedding operations at a fixed resolution, significantly reducing the computational complexity of embedding. Simultaneously, upsampling and overlay methods ensure the watermark's imperceptibility. The advantages of this embedding method are: firstly, high computational efficiency. For example, watermark embedding is performed at a low resolution (e.g., 256×256) instead of the original high resolution (e.g., 1920×1080), drastically reducing computation and significantly improving processing speed. Secondly, strong model versatility. For example, due to the fixed input size, the neural network watermark embedder model can operate independently of the original video resolution, eliminating the need to train multiple models for different resolutions. Furthermore, good watermark imperceptibility. For example, watermark distortion, after upsampling, is superimposed onto the original frame; due to the smoothing effect of upsampling, the distortion transitions smoothly at the pixel level, making it difficult to detect. Moreover, good robustness is maintained. For example, watermark distortion adapts to image content; watermark information embedded at low resolution is still retained after upsampling and can still be extracted under compression, transcoding, and other processing.

[0121] In some embodiments, embedding a neural network digital watermark in the frames of the video file to be published further includes employing a time watermark propagation strategy: performing a watermark embedding operation on one frame in every k frames to generate watermark distortion; and propagating the watermark distortion to the adjacent k-1 frames.

[0122] In this implementation, the server does not need to perform a complete watermark embedding calculation for every frame. It only needs to embed watermarks into some key frames and then copy or interpolate the watermark distortion to adjacent frames, which significantly improves the embedding throughput while maintaining robustness.

[0123] The above implementation embeds a digital watermark into the frames of the video file to be published, employing a time-based watermark propagation strategy to further reduce embedding computation costs and improve processing throughput. For example, firstly, the time-based watermark propagation parameter k can be pre-set according to the needs of the actual application scenario. k is a small integer representing the execution cycle of the watermark embedding operation, i.e., only one frame in every k frames undergoes a complete watermark embedding calculation. The value of k can range from 2 to 8; for example, k can be 2, 4, or 8. The value of k determines the balance between computational efficiency and robustness. The smaller the k value, the higher the execution frequency of the embedding operation, the denser the watermark distribution on the time axis, and the better the robustness, but the computational overhead is relatively large. The larger the k value, the lower the execution frequency of the embedding operation, and the higher the computational efficiency, but the sparser the watermark distribution on the time axis, which may slightly affect the robustness on the time axis.

[0124] Then, a complete watermark embedding operation can be performed on one frame in each of the k frames, generating watermark distortion. For example, the server groups the video sequence into groups of k frames. Within each group containing k frames, one frame (e.g., the first frame of each group, or a keyframe of each group) is selected as the anchor frame, and a complete watermark embedding operation is performed on that anchor frame. The complete watermark embedding operation includes: scaling the anchor frame to a preset resolution, invoking a neural network watermark embedder to embed a binary message in the pixel domain of the scaled frame, and generating watermark distortion. The generated watermark distortion is related to the size and content of the anchor frame, reflecting the amount of modification required in the pixel domain for watermark embedding. This watermark distortion will subsequently propagate to other frames within the group.

[0125] The generated watermark distortion can then be propagated to adjacent k-1 frames. For other k-1 frames (i.e., non-anchor frames) within the same group, the server does not re-perform the complete watermark embedding calculation, but instead reuses the watermark distortion generated by the anchor frame. Specifically, the watermark distortion generated by the anchor frame is directly copied or applied to other frames within the group via interpolation. Propagation methods can include direct copying or interpolation propagation. Direct copying involves directly copying the watermark distortion of the anchor frame to other frames within the group, with each non-anchor frame using the same watermark distortion. This method is simple to implement and has minimal computational overhead. Interpolation propagation involves interpolating the watermark distortion along the timeline, calculating the watermark distortion of intermediate frames based on the watermark distortion of adjacent anchor frames. For example, linear interpolation of the watermark distortion of two consecutive anchor frames can be used to smoothly transition the watermark along the timeline, further improving visual quality.

[0126] Finally, the watermark distortion after propagation can be upsampled and superimposed on the corresponding frame. For each frame in the group (including the anchor frame and other frames obtained from propagation), the corresponding watermark distortion is upsampled back to the original resolution and superimposed on the original frame with a preset intensity coefficient α to obtain the final watermarked frame.

[0127] With the above-mentioned time watermark propagation strategy, the server does not need to perform a complete watermark embedding calculation for every frame. It only needs to embed watermarks into some key frames and then copy or interpolate the watermark distortion to adjacent frames, which significantly improves the embedding throughput while maintaining robustness.

[0128] For example, when k=4, the server only needs to perform a complete embedding computation (including scaling, neural network inference, upsampling, etc.) on one of every four frames, while the other three frames only need to perform watermark distortion copying (or interpolation) and overlay operations. This improves the embedding speed by nearly 4 times and significantly reduces computational resource consumption, making it particularly suitable for real-time processing scenarios of high frame rate videos.

[0129] This time-based watermark propagation strategy significantly improves computational efficiency; for example, the embedding computation is reduced from once per frame to once per k frames, resulting in a processing speed increase of approximately k times. Using this strategy also maintains good visual quality. For instance, since adjacent frames have high content similarity, reusing the same watermark distortion does not introduce noticeable visual differences. Furthermore, robustness is generally well maintained; for example, because adjacent frames exhibit similar behavior during compression and transcoding, the propagated watermark distortion can still be stably extracted across frames. In addition, this strategy is highly scalable; for example, the value of k can be flexibly adjusted according to actual computing power requirements and robustness requirements to adapt to different deployment environments.

[0130] S403: Distribute the watermarked video file to the broadcast control terminal so that the broadcast control terminal can execute the above-described video playback control method.

[0131] In this step, the server distributes the watermarked video file generated by S402 to the broadcast control terminal, enabling the terminal to execute the aforementioned video playback control method. Distribution methods can include various forms such as content delivery network push, on-demand distribution, and multicast broadcast. The server can also record distribution logs during the distribution process, including distribution time, terminal identifier, and content identifier, for subsequent auditing and traceability.

[0132] This embodiment integrates a neural network watermark verifier into the terminal playback link to achieve real-time verification of playback frames. The verifier possesses dual capabilities: watermark payload recovery and malicious editing detection. Upon verification failure or detection of malicious editing, it immediately executes broadcast control blocking actions (such as pausing, black screen, or switching to a backup stream) and reports the failure information. Compared to traditional watermarking schemes that are only used for post-event tracing or offline detection, this achieves a real-time security closed loop of detection and blocking, enabling interception of abnormal content before it is broadcast.

[0133] Specifically, this disclosure achieves a closed-loop verifiable review-playback process. By embedding a watermark after server-center review and enforcing verification on the terminal side, the content played on the terminal is more strongly bound to the approved version, reducing the risk of unreviewed content being replaced. Furthermore, this disclosure simultaneously satisfies the requirements of benign robustness and malicious sensitivity. Through a dual-head neural network watermarking verifier structure and its accompanying training mechanism, it maintains a high pass rate for benign modifications such as transcoding and compression, while malicious editing such as overlapping characters, overlapping images, and cropping is detected and triggers failure. Moreover, it implements a real-time execution blocking mechanism on the terminal side. Through a lightweight neural network watermarking verifier and a sliding window fault-tolerance strategy, it achieves low latency and a controllable false alarm rate suitable for engineering deployment, realizing a stable closed-loop playback control system where failure results in interruption / blocking.

[0134] Reference Figure 6The diagram shown is a schematic representation of a video playback control device 600 according to an embodiment of this disclosure. It can be deployed on a broadcast control terminal and includes: The acquisition module 61 is used to acquire watermarked video files distributed locally. The watermarked video files are generated and distributed by the server after embedding digital watermarks into approved video files. The neural network watermark embedder is obtained by generating training video data based on benign transformations and malicious editing through dynamic data augmentation, and then co-training it with the neural network watermark verifier on the broadcast control terminal side by optimizing the joint loss function. The watermark embedded by the neural network watermark embedder is invisible to the naked eye, robust to preset benign transformations and can be successfully recovered by the neural network watermark verifier, and sensitive to malicious editing and can be detected by the neural network watermark verifier. Verification module 62 is used to call a neural network watermark verifier to verify the playback frame in the video playback processing link and obtain the verification result; the neural network watermark verifier includes a watermark payload recovery unit and a malicious editing detection unit, the watermark payload recovery unit is used to recover the watermark payload from the playback frame, and the malicious editing detection unit is used to detect whether the playback frame has been maliciously edited; the verification result includes watermark verification result and malicious editing detection result. The broadcast control module 63 is used to continue video playback if the verification result indicates that the verification is successful; and to execute a broadcast control blocking action and report failure information if the verification result indicates that the verification is unsuccessful.

[0135] Reference Figure 7 The diagram shown is a schematic of another video playback control device 700 according to an embodiment of this disclosure, which can be deployed on a server and includes: The acquisition module 71 is used to acquire the video file to be published, perform content review on the video file to be published, and output the review conclusion. The watermark embedding module 72 is used to embed a neural network digital watermark into the frames of the video file to be published if the review conclusion indicates that the review is passed, thereby generating a watermarked video file. The sending module 73 is used to distribute the watermarked video file to the broadcast control terminal so that the broadcast control terminal can execute the video playback control method described in any of the above embodiments.

[0136] The specific implementation of each of the above modules corresponds to the implementation process of the aforementioned method embodiments. For details, please refer to the relevant content in the aforementioned method embodiments, which will not be repeated here.

[0137] Reference Figure 8 The diagram shown is a schematic representation of the structure of a computer device 800 according to an exemplary embodiment of this disclosure, comprising: The processor 810, memory 820, and bus 830 are included. The memory 820 is used to store execution instructions and includes main memory 821 and external memory 822. The main memory 821, also known as internal memory, is used to temporarily store the operation data in the processor 810 and the data exchanged with external memory 822 such as hard disk. The processor 810 exchanges data with external memory 822 through main memory 821.

[0138] In this embodiment, the memory 820 is specifically used to store application code that executes the scheme of this disclosure, and its execution is controlled by the processor 810. That is, when the computer device 800 is running, the processor 810 communicates with the memory 820 through the bus 830, or the processor 810 communicates with the memory 820 through other means, so that the processor 810 executes the application code stored in the memory 820, and then executes the steps of the video playback control method described in any of the foregoing embodiments.

[0139] The memory 820 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0140] Processor 810 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0141] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the video playback control method described in any of the above embodiments.

[0142] Computer-readable storage media can be any usable medium accessible to a computer or a data storage device such as a server or data center that integrates one or more usable media. Usable media can be magnetic media, such as hard disks, floppy disks, and magnetic tapes; optical media, such as DVD-ROM, DVD-RAM, DVD-RW, DVD+RW, CD-ROM, CD-R, CD-RW, MO (magneto-optical) storage media; and semiconductor storage media, such as flash memory, EEPROM, Dynamic Random Access Memory (DRAM), and Static Random Access Memory (SRAM).

[0143] The computer program can be written in various computer programming languages, including but not limited to C, C++, Python, and custom messages and services under the ROS framework. When the computer program is executed by the processor, it implements the various steps of the video playback control method in the embodiments of this disclosure.

[0144] This disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of the video playback control method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0145] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0147] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0149] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0150] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A video playback control method, characterized in that, Applied to a broadcast control terminal, the method includes: The system acquires and distributes watermarked video files to the local machine. These watermarked video files are generated and distributed by the server after embedding a digital watermark into approved video files using a neural network watermark embedder. The neural network watermark embedder is obtained by generating training video data based on benign transformations and malicious editing through dynamic data augmentation, and then co-training it with a neural network watermark verifier on the broadcast control terminal side by optimizing the joint loss function. The watermark embedded by the neural network watermark embedder is invisible to the naked eye, robust to preset benign transformations and can be successfully recovered by the neural network watermark verifier, and sensitive to malicious editing and can be detected by the neural network watermark verifier. In the video playback processing chain, the neural network watermark verifier is invoked to verify the playback frame and obtain the verification result. The neural network watermark verifier includes a watermark payload recovery unit and a malicious editing detection unit. The watermark payload recovery unit is used to recover the watermark payload from the playback frame, and the malicious editing detection unit is used to detect whether the playback frame has been maliciously edited. The verification result includes the watermark verification result and the malicious editing detection result. If the verification result indicates that the verification is successful, continue video playback; if the verification result indicates that the verification is unsuccessful, execute the broadcast control blocking action and report the failure information.

2. The method according to claim 1, characterized in that, In the video playback processing chain, the neural network watermark verifier is invoked to verify the playback frames, including: After video frame decoding and before rendering and display, the neural network watermark verifier is called frame by frame for verification; or, a frame is sampled every preset number of frames, and the neural network watermark verifier is called for verification.

3. The method according to claim 1 or 2, characterized in that, The neural network watermark verifier is invoked to verify the playback frame, and the verification results are obtained, including: The playback frame is input into the neural network watermark verifier to obtain the watermark payload output by the watermark payload recovery unit and the tampering heatmap or tampering score output by the malicious editing detection unit. If the tampered score is greater than a preset tampered score threshold, or the maximum value of the tampered heatmap is greater than a preset maximum value threshold, or the average value of the tampered heatmap is greater than a preset average value threshold, then malicious editing is detected. If the confidence level, bit error rate, or p-value of the watermark payload does not meet the corresponding threshold, the watermark verification is deemed to have failed.

4. The method according to claim 1, characterized in that, If the verification result indicates that the verification failed, a broadcast control blocking action is executed and a failure message is reported, including: Obtain the verification results of multiple playback frames within a preset time window, and count the number of frames that were determined to be maliciously edited or failed to verify the watermark within the preset time window. If the statistical result exceeds the preset threshold, and the verification result indicates that the verification failed, then the broadcast control blocking action is triggered and the failure information is reported.

5. The method according to claim 1, characterized in that, The execution of the broadcast control blocking action and the reporting of failure information include: Execute a broadcast control blocking action, which includes at least one of the following: pausing playback, blacking out the screen, cutting off the video output signal, or switching to backup video content; Simultaneously extract the content identifier, timestamp, failure type, frame summary information and terminal device identifier of the playback frame, generate failure information and report it to the server; The failure types include watermark verification failure and / or detection of malicious editing.

6. The method according to claim 1, characterized in that, The neural network watermark validator and the server-side neural network watermark embedder are jointly trained in the following manner: The training video data is acquired, and the training enhancements are divided into a benign transformation set and a malicious editing set. The benign transformation set includes video encoding, video transcoding, image quality degradation, color or brightness adjustment, and minor geometric transformations. The malicious editing set includes text rendering overlay, image overlay, local repair, and cropping or trimming. The benign transformation samples are obtained by applying the transformations from the benign transformation set to the training video data. The transformations from the malicious editing set are applied to the training video data, and pixel-level mask labels are generated to obtain malicious editing samples and their corresponding pixel-level mask labels; Based on the benign transformation samples and the malicious editing samples and their corresponding pixel-level mask labels, the neural network watermark verifier and the neural network watermark embedder are jointly trained.

7. The method according to claim 6, characterized in that, The step of jointly training the neural network watermark verifier based on the benignly transformed samples and the maliciously edited samples and their corresponding pixel-level mask labels includes: Based on the benign transformation samples, the watermark payload recovery loss for the neural network watermark verifier and the neural network watermark embedder, the imperceptibility constraint loss for the neural network watermark embedder, and the false alarm suppression loss for the neural network watermark verifier are calculated. The false alarm suppression loss is used to constrain the tampering heatmap output by the malicious editing detection unit under benign transformation to be close to all zero. Based on the maliciously edited sample and the corresponding pixel-level mask label, calculate the tamper detection loss of the malicious editing detection unit for the neural network watermark verifier; The watermark payload recovery loss, the imperceptibility constraint loss, the false alarm suppression loss, and the tamper detection loss are weighted and summed to obtain a joint loss function. The neural network watermark verifier and the neural network watermark embedder are jointly trained by optimizing the joint loss function.

8. The method according to claim 7, characterized in that, The neural network watermark validator and the neural network watermark embedder are jointly trained by optimizing the joint loss function, including: Perform the first stage of training: using image data as training samples, jointly train the neural network watermark embedder and the neural network watermark verifier, and optimize the watermark payload recovery loss and imperceptible constraint loss in the joint loss function; The second phase of training is performed: image data and video data are used together as training samples. The training is carried out using benign transformations related to video compression and time watermark propagation strategies. Malicious edited samples and their pixel-level mask labels are also used for training. The tamper detection loss and false alarm suppression loss in the joint loss function are optimized. Perform the third stage of training: freeze the network parameters of the neural network watermark embedder, fine-tune the neural network watermark verifier, and optimize the tamper detection loss and false alarm suppression loss in the joint loss function.

9. A video playback control method, characterized in that, Applied to a server, the method includes: Obtain the video file to be published, conduct content review on the video file to be published, and output the review conclusion; If the review conclusion indicates that the review is passed, the neural network watermark embedder is invoked to embed a neural network digital watermark in the frames of the video file to be published, thereby generating a watermarked video file. The watermarked video file is distributed to the broadcast control terminal so that the broadcast control terminal can execute the video playback control method according to any one of claims 1 to 8.

10. The method according to claim 9, characterized in that, The process of embedding a neural network digital watermark into the frames of the video file to be published includes: In response to the review conclusion indicating that the review has been passed, the content identifier, policy version number and review time in the review conclusion are extracted and encoded into a binary message; The binary message is embedded as watermark information into the frames of the video file to be published, so that the watermarked video file carries the unique identifier of the review conclusion.

11. The method according to claim 9, characterized in that, The process of embedding a neural network digital watermark into the frames of the video file to be published includes: The frames of the video file to be published are scaled to a preset resolution to obtain scaled frames; The neural network watermark embedder is invoked to embed binary messages in the pixel domain of the scaled frame, generating watermark distortion. The watermark distortion is upsampled back to the original resolution and superimposed onto the original frame with a preset intensity coefficient to obtain a watermarked video file.

12. The method according to claim 11, characterized in that, The method of embedding a neural network digital watermark in the frames of the video file to be published also includes employing a time-based watermark propagation strategy. Perform a watermark embedding operation on one frame in each k-frame to generate watermark distortion; The watermark distortion is propagated to the adjacent k-1 frames.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions to perform the video playback control method as described in any one of claims 1-8 or 9-12.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the video playback control method as described in any one of claims 1-8 or 9-12.

Images