Multi-modal traceability method for generating penetrating and building videos for AI (Artificial Intelligence)

Abstract

The invention discloses a multi-modal traceability method for generating a penetrating and overlapping video for AI, and relates to the technical field of traceability management, and the method comprises the steps: deploying a multi-modal collection node comprising a video image collection point, an audio data collection point and a subtitle text collection point in an AI generation penetrating and overlapping video generation link; video frames, audio streams and subtitle data between a generation output end and a coding output end are collected in real time, an original video data set is constructed, standardization processing of video size, frame rate, color space and the like is executed based on edge tracing nodes, and a standard penetration video data set with a consistent structure and a unified format is output. The processing flow not only makes up for the defect that the existing AI video generation lacks chain acquisition record and structure uniformity, but also enhances the stability and reproducibility of subsequent watermark embedding and model fingerprint extraction by accurately setting the acquisition position and format conversion logic, and constructs a uniform input standard for multi-modal watermark and fingerprint verification.

Description

Technical Field

[0001] This invention relates to the field of traceability management technology, specifically a multimodal traceability method for AI-generated outfit videos. Background Technology

[0002] With the rapid development of AI-generated content technology, multimodal content such as images, audio, and video generated based on deep learning models has been widely applied in social media, entertainment media, advertising, and e-commerce, forming an intelligent creation ecosystem centered on content generation. Particularly in short video platforms and digital fashion scenarios, image generation models and video generation frameworks can be used to automatically generate fashion-related video content that includes character movements, clothing styles, and scene transitions. These AI-generated fashion videos not only possess highly human-like expressiveness but can also quickly generate short video materials that conform to fashion trends based on user-input style keywords or prompts, greatly improving content production efficiency and commercial conversion rates. Currently, AI-generated fashion videos have become a crucial supporting technology for applications such as fashion content marketing, e-commerce fashion displays, and virtual model try-on.

[0003] While AI-generated fashion videos currently demonstrate excellent performance in applications, they still face significant technical bottlenecks in areas such as originality verification, copyright confirmation, and determination of responsibility for generation. Firstly, due to the high malleability and non-uniqueness of AI-generated content, traditional visual similarity detection and image fingerprinting methods struggle to determine whether a video is original, edited, or generated from multiple sources. Secondly, mainstream video platforms lack a comprehensive traceability mechanism for AI-generated videos, making it difficult to trace key information such as the original generation model, generation parameters, and the generating entity if content is altered, source-changed, or involved in disputes after publication. Furthermore, existing digital watermarking technologies are mostly single-modal embedding methods, which are insufficient for comprehensive labeling and multi-angle verification in multimodal content like fashion videos that integrate visual actions, audio narration, and subtitles. This results in poor traceability stability, susceptibility to tampering and circumvention, and broken chains of evidence.

[0004] The root cause of these problems lies in the fact that AI-generated fashion videos typically lack key information collection nodes and structured registration mechanisms for tracing their origins. This makes it impossible to associate them with the original model identifier, generation parameter configuration, timestamp, and the identity of the original controller after generation. Furthermore, traditional video watermarks often use image frame embedding, lacking the ability to jointly mark audio and subtitle channels, making them prone to failure under complex editing operations and lacking robustness. In addition, due to the lack of trusted storage and on-chain verification channels, even if some watermark information exists during content generation, it is difficult to prevent subsequent tampering and forgery. Once such AI-generated fashion videos are edited and published or forged as original material without authorization, it will not only cause infringement disputes and brand damage, but may also lead to serious consequences such as public misleading, abuse of commercial resources, and increased compliance risks. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multimodal tracing method for AI-generated outfit videos, solving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multimodal source tracing method for AI-generated outfit videos, comprising the following steps:

[0007] S1. Deploy multimodal acquisition nodes on the video generation chain to collect the original video dataset of AI-generated outfit videos in real time, and upload the original video dataset to the edge tracing node; perform normalization processing on the original video dataset in the edge tracing node to obtain a normalized outfit video dataset;

[0008] S2. Based on the standardized dressing video dataset, generate a registered watermark reference sequence WB and register the registered watermark reference sequence WB in the traceability record database;

[0009] S3. Extract the watermark bit sequence WA based on the video data to be verified, and then calculate the error rate with the registered watermark reference sequence WB to obtain the watermark bit error rate BER. Then compare the watermark bit error rate BER with the preset watermark interval threshold to determine the current watermark verification status and trigger fingerprint depth verification processing.

[0010] S4. When fingerprint depth verification is triggered, perform fingerprint depth verification, extract the model fingerprint vector VA corresponding to the video to be verified, and calculate the similarity value SIM with the registered model fingerprint vector VB.

[0011] S5. Then, preset the similarity threshold Sth and the similarity value SIM to evaluate the watermark similarity, and execute the corresponding control actions based on the evaluation results.

[0012] Preferably, S1 includes S11;

[0013] S11. In the outfit video generation system, a multimodal acquisition node is set between the generation output end and the encoding output end on the AI-generated outfit video generation link. Through the multimodal acquisition node, multimodal data is collected during the generation and verification process of AI-generated outfit videos to obtain the original video dataset. The original video dataset is then uploaded to the edge tracing node through the edge communication interface.

[0014] The multimodal acquisition nodes include video image acquisition points, audio data acquisition points, and text data acquisition points;

[0015] The original video dataset includes a set of joint frames F, an optical flow field set O, a saliency map set M, a DCT coefficient set C, an audio spectrogram set A, and a subtitle timeline set T.

[0016] Preferably, S1 further includes S12;

[0017] S12. In the edge tracing node, the original video dataset is normalized to obtain a standardized outfit video dataset.

[0018] The normalization process includes video normalization, audio normalization, and subtitle text normalization.

[0019] The video normalization process scales the video frames to a preset standard size of 4K using data collected from video capture points, then resamples the video frame rate to a uniform frame rate of 60fps, and finally converts the original image data into a standard color space of RGBBT.709.

[0020] The audio normalization process involves performing sampling rate unification processing to unify the audio sampling rate to a standard value of 44100Hz, and then performing channel format normalization processing to downmix multi-channel audio into mono or stereo format.

[0021] The subtitle text normalization process extracts the start and end time information of the subtitles to form a subtitle timeline set T; based on the video frame timeline and audio timeline, the subtitle display nodes are calibrated.

[0022] Preferably, S2 includes S21;

[0023] S21. Based on the standardized outfit video dataset, watermark embedding processing is performed on the main clothing area and dynamic area in the outfit video. The watermark embedding processing is performed by executing a multimodal watermark embedding strategy set. Based on the execution result of the multimodal watermark embedding strategy set, the embedded watermark bit sequence is extracted from the standard frame position and frequency domain position during the embedding process, a registration watermark reference sequence WB is generated, and a registration operation is performed.

[0024] The multimodal watermarking embedding strategy set includes region selection strategy K1, time-domain diffusion strategy K2, frequency-domain embedding strategy K3, dynamic adaptive strategy K4, and error correction coding strategy K5;

[0025] The registration operation combines the registration watermark reference sequence WB with the corresponding outfit video's generation model identifier, generation parameter summary, generation timestamp, and generation metadata to form a traceability data summary;

[0026] The trusted execution environment at the generation end performs digital signatures on the registered watermark reference sequence WB and the source data digest;

[0027] The signed registration watermark reference sequence WB and the generated metadata are written into the registration traceability record set as registration traceability records and simultaneously written into the blockchain ledger.

[0028] Preferably, S3 includes S31;

[0029] S31. Based on the AI-generated outfit video to be verified, extract the watermark extraction sequence WA using the same multimodal acquisition node; then calculate the watermark bit error rate BER of the current AI-generated outfit video by comparing the watermark extraction sequence WA with the registered watermark reference sequence WB.

[0030] Preferably, S3 further includes S32;

[0031] S32. In the edge tracing node, two watermark interval thresholds related to watermark verification are pre-set. The watermark interval thresholds include the watermark passing threshold Bth1 and the strong verification trigger threshold Bth2. Then, the real-time watermark bit error rate BER is compared with the two watermark interval thresholds to determine the current watermark verification status and trigger model fingerprint analysis and blockchain trust verification. The specific comparison content is as follows:

[0032] When the watermark bit error rate BER ≤ watermark passing threshold Bth1, the watermark consistency of the current AI-generated outfit video to be verified is determined to meet the passing condition, the watermark passing status result is output, and the AI-generated outfit video to be verified is marked as a watermark verification passed video.

[0033] When the watermark passes the threshold Bth1 < watermark bit error rate BER ≤ strong verification trigger threshold Bth2, it is determined that the current AI-generated outfit video to be verified is in an uncertain state of watermark verification, the watermark gray area state result is output, and fingerprint depth verification processing is triggered.

[0034] When the watermark bit error rate BER is greater than the strong verification trigger threshold Bth2, the watermark consistency of the current AI-generated outfit video to be verified is deemed insufficient, the watermark verification failure result is output, and fingerprint deep verification processing is forcibly triggered.

[0035] Preferably, S4 includes S41;

[0036] S41. When fingerprint depth verification is triggered, extract the model fingerprint vector VA of the outfit video generated by the AI ​​to be verified and the registered model fingerprint vector VB; the specific extraction content is as follows;

[0037] Based on the standardized outfit video dataset corresponding to AI-generated outfit videos, key frames representing the outfit scene content are extracted from the key frame set F;

[0038] Keyframes representing the content of the outfit image are input into a pre-defined model fingerprint feature extraction network to extract the intermediate layer feature responses of the model fingerprint feature extraction network.

[0039] Feature aggregation processing is performed on the intermediate layer feature responses corresponding to multiple representative keyframes, and the aggregation results are normalized and dimensionally constrained to generate the fingerprint vector VA of the current video model.

[0040] In the outfit video generation stage, based on the same network structure, key frame selection rules and feature processing flow as the model fingerprint feature extraction network, the generated outfit video is subjected to model fingerprint extraction processing to generate the registered model fingerprint vector VB.

[0041] The registration model fingerprint vector VB, along with the corresponding outfit video identifier information, watermark reference sequence, and generated metadata, are written into the registration traceability record set.

[0042] At the same time, the unique video identifier ID of the outfit video to be verified is read from the registration and traceability record set, and the registration model fingerprint vector VB, which was registered synchronously when the video was generated, is read through the unique video identifier ID. The registration model fingerprint vector VB is a vector generated by the generator based on the same features extracted by the network and rules as the model fingerprint vector VA, and is encrypted, signed and stored in the central database or blockchain node.

[0043] Preferably, S4 further includes S42;

[0044] S42. Based on the model fingerprint vector VA corresponding to the obtained video to be verified and the registered model fingerprint vector VB, perform cosine similarity calculation, output similarity value SIM, and quantitatively analyze the similarity between the AI-generated outfit video and the level model.

[0045] Preferably, S5 includes S51;

[0046] S51. Using a pre-built video generation model fingerprint sample library, statistically analyze the similarity value SIM interval of a large number of registered AI-generated outfit videos. Based on the lower bound of the interval where the statistical distribution is at least 95% confidence level, set the similarity threshold Sth for the model fingerprint. Then, evaluate the watermark similarity between the real-time acquired model fingerprint vector VA and the registered model fingerprint vector VB and the similarity threshold Sth.

[0047] Simultaneously, the generated model fingerprint vector VA, the registered watermark reference sequence WB, and the generation timestamp extracted from the AI-generated outfit video to be verified are performed on-chain. The on-chain verification is achieved by constructing a traceability hash value and submitting it to the on-chain database node. The database node then performs parallel verification on the hash matching, the validity of the registered generator signature, and the logical order of the timestamp. If all three verifications pass, the on-chain verification is reported as passed; if any one verification fails, the on-chain verification is reported as failed.

[0048] The specific assessment content is as follows:

[0049] When the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is greater than or equal to the similarity threshold Sth, and the verification is passed, it is determined that the two are generated from the same source, and the current AI-generated outfit video is marked as having passed the similarity verification.

[0050] If the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is less than the similarity threshold Sth, and the verification is passed, it is determined that there is tampering, and the current AI-generated outfit video is marked as a similarity verification failure.

[0051] When on-chain verification fails, the registration is deemed unverifiable, and the current AI-generated outfit video is marked as having an unknown source.

[0052] Preferably, S5 further includes S52;

[0053] S52. Based on the judgment results of watermark similarity assessment and on-chain verification, execute the control actions corresponding to the verification status, specifically including:

[0054] When the similarity verification of the judgment result is passed, the corresponding AI-generated outfit video will be marked as publicly shared and the video platform will be allowed to open the dissemination permission;

[0055] When the similarity verification fails, the content source tracing manual review process is triggered.

[0056] When the on-chain verification result is unsuccessful, the corresponding AI-generated outfit video will be directly identified as a video of unknown origin, and risk labeling will be applied to the AI-generated outfit video to restrict its dissemination and use, triggering a manual review process for content tracing.

[0057] This invention provides a multimodal source tracing method for AI-generated outfit videos. It has the following beneficial effects:

[0058] (1) This method deploys multimodal acquisition nodes, including video frame acquisition points, audio data acquisition points, and subtitle text acquisition points, in the AI-generated outfit video generation chain. It collects video frames, audio streams, and subtitle data between the generation output and encoding output in real time, constructing an original video dataset containing a set of joint frames F, an optical flow field set O, a saliency map set M, a DCT coefficient set C, an audio spectrogram set A, and a subtitle timeline set T. Based on edge tracing nodes, it performs normalization processing on video size, frame rate, and color space, outputting a standardized outfit video dataset with consistent structure and format. This processing flow not only compensates for the lack of on-chain acquisition records and structural uniformity in existing AI-generated videos, but also enhances the stability and reproducibility of subsequent watermark embedding and model fingerprint extraction by precisely setting the acquisition position and format conversion logic. It establishes a unified input standard for multimodal watermarking and fingerprint verification, thereby achieving a high-fidelity and high-stability video tracing data structure.

[0059] (2) This method constructs a multimodal watermark embedding strategy set based on five strategies—saliency region localization, temporal diffusion, frequency domain embedding, dynamic adaptation, and error correction coding—on a standardized clothing video dataset. Watermark information is embedded into the clothing areas of the clothing videos to generate a registration watermark reference sequence WB, which is then registered in the registration traceability record set. During the verification phase, the watermark extraction sequence WA is extracted, the bit error rate BER is calculated, and a threshold range is set to distinguish between pass, gray zone, and failure states, triggering subsequent fingerprint depth verification. This verification uses a convolutional neural network (CNN) to extract intermediate layer features from keyframes, generating a model fingerprint vector VA. A SIM similarity calculation is performed between this model fingerprint vector VA and the registered model fingerprint vector VB, and a similarity threshold Sth is set using a confidence interval. Simultaneously, on-chain verification is performed based on the blockchain to ensure the authenticity and immutability of the registered data. This mechanism has higher anti-editing capabilities and fault tolerance than existing "single watermark verification" or "image hash verification" technologies, ensuring stable source verification even after complex processing such as editing, compression, and encapsulation. It also effectively identifies counterfeit content, improving traceability security and legal validity.

[0060] (3) This method combines a multimodal acquisition and normalization processing mechanism based on the generation link with a dual verification mechanism that integrates watermarking and model fingerprinting. This invention constructs an end-to-end verifiable, cross-modal trusted registration, and blockchain-supported full lifecycle traceability and ownership confirmation system for AI-generated fashion videos. This system has multi-layer information acquisition, change perception, and consistency verification capabilities from the generation source, transmission path to the verification end. It can effectively bridge the trust gap between generated content and true ownership, and provide comprehensive technical support in anti-counterfeiting, ownership confirmation, supervision, and content responsibility attribution. It is particularly suitable for high-frequency generation, rapid dissemination, and copyright-sensitive business scenarios of fashion videos, and helps to build a healthy AIGC video content ecosystem and industry compliance infrastructure. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the steps of a multimodal source tracing method for AI-generated outfit videos according to the present invention;

[0062] Figure 2 This is a flowchart for video standardization processing. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Example 1

[0065] This invention provides a multimodal source tracing method for AI-generated outfit videos. Please refer to [link / reference]. Figure 1 This includes the following steps:

[0066] S1. Deploy multimodal acquisition nodes on the video generation chain to collect the original video dataset of AI-generated outfit videos in real time, and upload the original video dataset to the edge tracing node; perform normalization processing on the original video dataset in the edge tracing node to obtain a normalized outfit video dataset;

[0067] S2. Based on the standardized dressing video dataset, generate a registered watermark reference sequence WB and register the registered watermark reference sequence WB in the traceability record database;

[0068] S3. Extract the watermark bit sequence WA based on the video data to be verified, and then calculate the error rate with the registered watermark reference sequence WB to obtain the watermark bit error rate BER. Then compare the watermark bit error rate BER with the preset watermark interval threshold to determine the current watermark verification status and trigger fingerprint depth verification processing.

[0069] S4. When fingerprint depth verification is triggered, perform fingerprint depth verification, extract the model fingerprint vector VA corresponding to the video to be verified, and calculate the similarity value SIM with the registered model fingerprint vector VB.

[0070] S5. Then, preset the similarity threshold Sth and the similarity value SIM to evaluate the watermark similarity, and execute the corresponding control actions based on the evaluation results.

[0071] In this embodiment, the method deploys multimodal acquisition nodes in the AI-generated outfit video generation chain to synchronously collect raw data from modalities such as images, audio, and text from the source. This effectively solves the problems of information loss and difficulty in identifying tampering caused by video encoding, editing, or compression in traditional source tracing methods. By collecting data at the generation end, the originality and integrity of subsequent watermark embedding and model fingerprint extraction can be ensured, avoiding the ambiguity and misjudgment risks brought about by verification based solely on the content at the publishing end. The collected raw data undergoes normalization processing at the edge source tracing nodes, such as unifying resolution, frame rate, and color space, to ensure the consistency of video content in structural features under different platform distribution or post-processing, thereby improving the robustness of subsequent watermark extraction and model fingerprint comparison. Normalized video data is used to perform multimodal watermark embedding. The embedding strategy integrates saliency region selection, temporal diffusion, frequency domain embedding, and dynamic adjustment to achieve high robustness of the watermark under the premise of perceptual non-disruption. Error correction coding is used to enhance the extraction fault tolerance capability, ultimately forming a registered watermark reference sequence WB. This sequence, along with metadata such as the generation model identifier and generation parameters, is written into the registration traceability record set, with optional on-chain synchronization to enhance immutability. In the verification phase, the watermark bit sequence WA is extracted from the video to be verified and the watermark bit error rate BER is calculated by comparing it with the registered watermark reference sequence WB. BER=0 indicates complete consistency, while a higher BER indicates a greater difference. Two watermark interval thresholds, Bth1 and Bth2, are set to determine whether the video can directly pass verification or needs to enter the deep verification process, respectively. This method, by introducing a "trusted gray zone" setting, effectively avoids misidentification caused by a one-size-fits-all judgment, enhancing the system's adaptability and fault tolerance. When the watermark bit error rate (BER) exceeds Bth1 or Bth2, the system automatically triggers fingerprint depth verification. A pre-defined convolutional neural network (CNN) is used to extract model fingerprints from keyframes. After extracting intermediate layer feature vectors, aggregation and normalization are performed to obtain the model fingerprint vector VA for the video to be verified. During the generation phase, a registration fingerprint vector VB is generated using the same network and processing rules and has been written into the registration traceability record set. A similarity value (SIM) is calculated between VA and VB to evaluate the consistency of their generation models. The distribution of the similarity value SIM is statistically analyzed using a pre-built fingerprint sample library. A similarity threshold (Sth) is set at a 95% confidence level to determine whether two videos were generated from the same source. Finally, the similarity value SIM is compared with the similarity threshold Sth, and verification is performed in conjunction with on-chain hash registration information, signature validity, and timestamp order. If SIM ≥ similarity threshold Sth and on-chain verification passes, the video is determined to be trusted generated content and can be freely disseminated; if SIM < similarity threshold Sth but on-chain verification passes, it is marked as tampered content and enters manual review; if on-chain verification fails, it is marked as a video of unknown origin and its dissemination and use are restricted.This comprehensive implementation method forms a trusted traceability mechanism covering the entire process "from generation to governance," possessing four key capabilities: link-level data collection assurance, modality-level watermark embedding, feature-level deep fingerprint comparison, and strategy-level control linkage. It achieves closed-loop risk control and traceable management of AI-generated fashion videos from source, pathway, to the end user. This significantly improves the verifiability, security, and platform governance efficiency of AIGC content, demonstrating significant practical applications in addressing the proliferation of fake content, AI-generated content ownership verification, and deepfake detection.

[0072] Example 2

[0073] Please see Figure 1 and Figure 2 Specifically: S1 includes S11;

[0074] S11. In the outfit video generation system, a multimodal acquisition node is set between the generation output end and the encoding output end on the AI-generated outfit video generation link. Through the multimodal acquisition node, multimodal data is collected during the generation and verification process of AI-generated outfit videos to obtain the original video dataset. The original video dataset is then uploaded to the edge tracing node through the edge communication interface for subsequent tracing processing.

[0075] Multimodal acquisition nodes include video capture points, audio data capture points, and text data capture points;

[0076] The original video dataset includes a set of joint frames F, an optical flow field set O, a saliency map set M, a DCT coefficient set C, an audio spectrogram set A, and a subtitle timeline set T;

[0077] The video capture point is located between the AI-generated video module and the video encoding module. Specifically, the image frames output by the generated model are sent to the video capture point and then to the encoder (such as H.264 / H.265). The video frame listening module (which can be implemented as a "frame capture module" embedded in the GPU / edge computing node) completes the data capture, or it is read asynchronously through the frame buffer. The set of joint frames F, the set of optical flow fields O, the set of saliency maps M, and the set of DCT coefficients C are acquired in real time.

[0078] The audio data acquisition point is set between the speech generation module and the audio encapsulation module. That is, the TTS model outputs audio to the audio acquisition point, and then it is encapsulated into the audio track in the video stream (MP4 / AVI). The audio stream listener or the PCM buffer reading module based on the system pipeline is used for non-interference reading; the audio spectrogram set A is acquired in real time.

[0079] The text data acquisition point is set between the subtitle generation module and the video synthesis module. That is, the text generated by the LLM or prompt template is sent to the text acquisition point, and then the synthesized subtitles, overlaid tags, and text-to-speech are obtained directly from the output results of the text generation cache queue or language model calling interface, or extracted through the LLM output log monitoring module; the subtitle timeline set T is collected in real time.

[0080] The keyframe set F is used to extract keyframes representing the content of the outfit from consecutive video frames generated by AI, so as to reflect the representative static features of the outfit.

[0081] The optical flow field set O is used to calculate motion change information of the human body and clothing areas based on continuous video frames, so as to describe the dynamic change characteristics in the dressing video;

[0082] A set of saliency maps M is used to perform saliency analysis on video footage to locate the main clothing area in an outfit video;

[0083] The DCT coefficient set C is used to perform frequency domain transformation on video frames and extract the corresponding frequency domain feature parameters.

[0084] Audio spectrogram set A is used to perform spectral analysis on the audio data corresponding to the outfit video in order to obtain the audio frequency domain characteristics.

[0085] The subtitle timeline set T is used to record the time correspondence between subtitle text and video footage and audio data.

[0086] S1 also includes S12;

[0087] S12. In the edge tracing node, the original video dataset is normalized to obtain a standardized outfit video dataset.

[0088] Standardization processing includes video standardization processing, audio standardization processing, and subtitle text standardization processing;

[0089] Video normalization processing scales video frames to a preset standard size of 4K using data collected from video capture points to eliminate differences in encoding resolution across different platforms. Then, the video frame rate is resampled to a uniform frame rate of 60fps to maintain video rhythm consistency. Finally, the original image data is uniformly converted to the standard color space RGBBT.709 to facilitate DCT frequency domain processing and saliency map calculation.

[0090] Audio normalization processing involves performing sampling rate unification processing to unify the audio sampling rate to the standard value of 44100Hz, and then performing channel format normalization processing to downmix multi-channel audio into mono or stereo format for easier subsequent spectrum analysis.

[0091] The subtitle text standardization process extracts the start and end time information of the subtitles to form a subtitle timeline set T; based on the video frame timeline and audio timeline, the subtitle display nodes are calibrated to ensure that the text content is synchronized with the picture and audio.

[0092] In this embodiment, in the AI-generated outfit video generation chain, S11 deploys a multimodal acquisition node between the generation output end and the encoding output end. This enables the synchronous capture of original image, audio, and subtitle data, avoiding data distortion introduced by subsequent encoding and encapsulation processes. For example, if the acquisition node is set at the video publishing end, the video frames compressed by H.264 will cause DCT frequency domain feature distortion, failing to accurately reflect the model-generated features. Therefore, this pre-acquisition at this location ensures the authenticity and integrity of the data source, avoiding analysis bias caused by "modal drift." Through the joint frame set F and the optical flow field set O, the dynamic information of the human body and clothing area can be clearly captured, avoiding the technical defects of traditional image frames that cannot fully reproduce the outfit movement due to occlusion or background interference. At the same time, by locking the main clothing area through the saliency map set M, the problem of watermark and fingerprint features falling into the background area and becoming ineffective is effectively reduced. The separate audio and subtitle acquisition points ensure precise synchronization between the audio spectrogram set A and the subtitle timeline set T from TTS output to encapsulation. This solves the audio-visual asynchrony problem caused by buffer delays or subtitle drift in the generated model, ensuring temporal consistency of the three-modal data in subsequent source tracing verification. Normalization in S12 addresses feature extraction errors caused by inconsistent video formats generated by different platforms or models. For example, inconsistent resolutions can lead to salient region localization errors, and inconsistent frame rates can cause dynamic feature dilution. Therefore, unifying video frames to 4K, 60fps and converting them to the standard color space RGBBT.709 effectively constructs a unified data feature space, making subsequent DCT frequency domain processing, model fingerprint extraction, and watermark embedding more stable and universal. The audio sampling rate is unified to 44100Hz because some TTS models have low sampling rate outputs (e.g., 22kHz), resulting in insufficient spectral line density in the spectrogram and affecting the accuracy of frequency domain feature extraction. Normalization ensures the comparability of spectrum analysis. Subtitle text normalization calibrates the start and end nodes of subtitles to address the issue of subtitle timeline drift that occurs in multiple output rounds of the generation model. By aligning with the audio and frame axes, a strong binding relationship between subtitles and the image is ensured, facilitating subsequent behavioral sequence reasoning or temporal integrity verification.

[0093] Example 3

[0094] Please see Figure 1 Specifically: S2 includes S21;

[0095] S21. Based on the standardized clothing video dataset, watermark embedding is performed on the main clothing area and dynamic area in the clothing video. The watermark embedding process is carried out by executing a multimodal watermark embedding strategy set. Based on the execution result of the multimodal watermark embedding strategy set, the embedded watermark bit sequence is extracted from the standard frame position and frequency domain position during the embedding process, a registration watermark reference sequence WB is generated, and the registration operation is performed.

[0096] The set of multimodal watermarking embedding strategies includes region selection strategy K1, time-domain diffusion strategy K2, frequency-domain embedding strategy K3, dynamic adaptive strategy K4, and error correction coding strategy K5;

[0097] The registration process combines the registration watermark reference sequence WB with the corresponding outfit video's generation model identifier, generation parameter summary, generation timestamp, and generation metadata to form a traceability data summary.

[0098] The trusted execution environment at the generation end performs digital signatures on the registered watermark reference sequence WB and the source data digest;

[0099] The signed registration watermark reference sequence WB and the generated metadata are written into the registration traceability record set as registration traceability records and simultaneously written into the blockchain ledger to improve anti-tampering capabilities.

[0100] Among them, based on the saliency map set M, the salient areas of clothing are located, and the region selection strategy K1 is executed to prioritize embedding watermark information in the edge areas of clothing texture and avoid face and background areas.

[0101] Based on the keyframe set F, a temporal diffusion strategy K2 is executed to periodically and repeatedly embed the same watermark bit sequence in multiple keyframes, thereby improving the robustness to editing operations.

[0102] Based on the DCT coefficient set C, the frequency domain embedding strategy K3 is executed to embed the watermark bits into the mid-frequency band position, thereby reducing the impact of compression on the watermark.

[0103] Based on the optical flow field set O, a dynamic adaptive strategy K4 is executed to dynamically adjust the embedding intensity, resulting in higher watermark stability in low motion regions.

[0104] Error correction coding strategy K5 is applied to the original watermark bit sequence, and the BCH coding algorithm is used to increase redundant information and enhance the error tolerance capability.

[0105] In this embodiment, the introduction of a multimodal watermark embedding strategy set during the implementation of S21 is to address the problems of traditional watermark embedding relying solely on image pixel domains and being easily destroyed by compression and editing operations. For example, if the watermark is randomly embedded into the entire image, it is easily lost during video transcoding or background blurring. By using the region selection strategy of K1, the watermark is explicitly embedded only in the edge area of ​​the clothing texture, avoiding the facial area, which prevents the watermark from interfering with the visual focus and improves concealment; because the clothing texture is highly complex, it is more suitable for loading concealed information.

[0106] K2's temporal diffusion strategy avoids "editing corruption"; if the watermark is only embedded in a few frames, editing will directly destroy the verification basis; while periodic diffusion embedding can ensure that some watermark fragments are retained, and it is still verifiable even if it is edited.

[0107] K3 avoids the problem of low-frequency components being extremely sensitive to video compression by using DCT mid-frequency embedding, and also avoids noise interference in the high-frequency region. Choosing mid-frequency embedding is a "frequency domain steady-state region" strategy that balances robustness and concealment. This setting shows a high survival rate under actual compression (such as H.264).

[0108] K4's dynamic adaptive strategy addresses the phenomenon that watermark displacement or deformation can occur in areas of intense motion in videos. By using optical flow field information, it automatically reduces the embedding intensity in high-motion areas, avoiding the "dynamic erasure" phenomenon and making the watermark more stable and persistent in static areas, thus enhancing the recognition rate in motion scenes.

[0109] K5 introduces the BCH error-correcting coding algorithm to address potential bit errors during actual extraction. Especially after multiple compressions or filtering processes, watermarks may exhibit unstructured bit flips. Without an error-correcting mechanism, the entire watermark would become unrecognizable. BCH coding possesses strong redundancy capabilities, effectively recovering even with 20% bit errors, significantly improving the watermark's anti-interference performance.

[0110] Furthermore, the entire embedding process is synchronously signed and registered by a trusted execution environment and can be recorded on the blockchain, ensuring that the watermark and generated metadata have "identity binding" capabilities. This operation prevents forgers from tampering with the generated parameters to evade responsibility later, playing a true "responsibility anchoring" role. Overall, this step, through multi-strategy collaboration and multi-dimensional spatial and temporal embedding, constructs a watermark structure with high concealment, high stability, verifiability, and registrability, providing a strong data foundation for subsequent traceability and verification.

[0111] Example 4

[0112] Please see Figure 1 Specifically: S3 includes S31;

[0113] S31. Based on the AI-generated outfit video to be verified, extract the watermark extraction sequence WA using the same multimodal acquisition node; then calculate the watermark bit error rate BER of the current AI-generated outfit video by comparing the watermark extraction sequence WA with the registered watermark reference sequence WB.

[0114] The watermark bit error rate (BER) is calculated and output using the following algorithm formula;

[0115] ;

[0116] In the formula, BER(WA, WB) represents the watermark bit error rate between the watermark extraction sequence WA and the registered watermark reference sequence WB, m represents the length of the watermark bit sequence, WA[j] represents the j-th bit in the watermark extraction sequence WA, and WB[j] represents the j-th bit in the registered watermark reference sequence WB. This represents the XOR operation;

[0117] Bit Error Rate (BER) is a fundamental evaluation formula derived from Information Theory and digital communication system theory. It is used to measure the degree of difference between two bit sequences and is a classic metric for measuring the bit error rate of signal transmission in communication systems.

[0118] In this formula, it is used to measure the consistency between the watermark extraction sequence WA of AI-generated outfit videos and the registered watermark reference sequence WB. This is a typical transplantation and application of the formula in the field of digital content security.

[0119] The BER value output by this formula can be regarded as: "the degree of logical difference between the extracted watermark and the original registered watermark";

[0120] Specifically, BER=0 indicates complete consistency, while the closer the BER is to 1, the worse the consistency. This bit error rate has extremely high sensitivity and stability, and is suitable for detecting the degree of damage to the watermark during the transmission, transcoding, editing, and splicing of clothing videos.

[0121] S3 also includes S32;

[0122] S32. In the edge tracing node, two watermark interval thresholds related to watermark verification are pre-set. The watermark interval thresholds include the watermark passing threshold Bth1 and the strong verification trigger threshold Bth2. Then, the real-time watermark bit error rate BER is compared with the two watermark interval thresholds to determine the current watermark verification status and trigger model fingerprint analysis and blockchain trust verification. The specific comparison content is as follows:

[0123] When the watermark bit error rate BER ≤ watermark passing threshold Bth1, the watermark consistency of the current AI-generated outfit video to be verified is determined to meet the passing condition, the watermark passing status result is output, and the AI-generated outfit video to be verified is marked as a watermark verification passed video.

[0124] When the watermark passes the threshold Bth1 < watermark bit error rate BER ≤ strong verification trigger threshold Bth2, it is determined that the current AI-generated outfit video to be verified is in an uncertain state of watermark verification, the watermark gray area state result is output, and fingerprint depth verification processing is triggered.

[0125] When the watermark bit error rate BER > the strong verification trigger threshold Bth2, it is determined that the watermark consistency of the current AI-generated outfit video to be verified is insufficient, the watermark verification failure result is output, and fingerprint deep verification processing is forcibly triggered.

[0126] The watermark interval threshold is determined by setting 100 training set samples, which include real generated videos and edited, spliced, and forged videos. The watermark bit error rate (BER) of all videos is extracted and analyzed. The maximum BER of real generated videos and the minimum BER of edited, spliced, and forged videos are set as the watermark passing threshold Bth1 and the strong verification trigger threshold Bth2, respectively.

[0127] In this embodiment, the method uses the watermark bit error rate (BER) in S31 as the criterion for determining the consistency between the extracted sequence WA and the registered sequence WB. This enables precise quantitative assessment of whether the watermark in the AI-generated outfit video has been tampered with, edited, or compressed. This is because traditional video similarity methods struggle to identify subtle deviations in hidden information, while BER, based on bitwise XOR calculations, can quickly generate difference feedback when minor tampering occurs, exhibiting high sensitivity and effectively capturing microscopic destructive behavior. In S32, a dual threshold mechanism is introduced, with the watermark passing through threshold Bth1 and a strong verification trigger threshold Bth2. This effectively avoids the "fuzzy false positive boundary" problem caused by traditional single thresholds. For example, if the system only sets a single pass standard, false positives or false negatives are highly likely when the video is in a blurred edge state (such as after slight transcoding). Setting a gray zone judgment interval allows for further verification of videos in a critical state, improving the overall judgment accuracy. In particular, by incorporating BER distribution statistics of real and fake videos into the training samples to set Bth1 and Bth2, the judgment criteria are ensured to originate from a data-driven, reliable distribution range, avoiding subjective bias caused by human intervention. This strategy is quantifiable and transferable. In summary, this implementation process, by constructing a watermark consistency detection mechanism centered on BER and supplementing it with a multi-threshold triggering strategy based on data statistics, achieves a complete closed loop from accurate detection to intelligent judgment. Physically, it forms a numerical expression and hierarchical judgment of the watermark structural integrity. When facing attacks such as editing, compression, and forgery, the system can stably and automatically perform source tracing verification, effectively improving the controllability and traceability of AI-generated video content.

[0128] Example 5

[0129] Please see Figure 1 Specifically: S4 includes S41;

[0130] S41. When fingerprint depth verification is triggered, extract the model fingerprint vector VA of the outfit video generated by the AI ​​to be verified and the registered model fingerprint vector VB; the specific extraction content is as follows;

[0131] Based on the standardized outfit video dataset corresponding to AI-generated outfit videos, key frames representing the outfit scene content are extracted from the key frame set F;

[0132] The keyframes representing the content of the outfit image are input into a pre-defined model fingerprint feature extraction network to extract the intermediate layer feature responses of the model fingerprint feature extraction network.

[0133] Feature aggregation processing is performed on the intermediate layer feature responses corresponding to multiple representative keyframes, and the aggregation results are normalized and dimensionally constrained to generate the fingerprint vector VA of the current video model.

[0134] In the outfit video generation stage, based on the same network structure, key frame selection rules and feature processing flow as the model fingerprint feature extraction network, the generated outfit video is subjected to model fingerprint extraction processing to generate the registered model fingerprint vector VB.

[0135] The registration model fingerprint vector VB, along with the corresponding outfit video identifier information, watermark reference sequence, and generated metadata, are written into the registration traceability record set.

[0136] At the same time, the unique video identifier ID of the outfit video to be verified is read from the registration and traceability record set, and the registration model fingerprint vector VB, which is registered synchronously when the video is generated, is read through the unique video identifier ID. The registration model fingerprint vector VB is a vector generated by the generator based on the same features extracted by the network and rules as the model fingerprint vector VA, and is encrypted, signed and stored in the central database or blockchain node.

[0137] To achieve in-depth source tracing and verification of AI-generated outfit videos, a model fingerprint vector is generated and registered based on a model fingerprint feature extraction network. The model fingerprint feature extraction network is constructed using a convolutional neural network (CNN), and features are extracted, aggregated, and normalized by combining a set of key frames. Finally, a structural vector representation strongly correlated with the video generation model is generated, which facilitates subsequent source tracing, matching, and forgery detection.

[0138] Input data construction:

[0139] In the outfit video generation system, AI-generated outfit videos undergo normalization processing to generate a standardized outfit video dataset. Based on this dataset, a keyframe set F={f1,f2,...,fn} is extracted, where n=10. The sampling rule is to extract one frame every 1 second, prioritizing frames containing the main movement area of ​​the clothing to enhance the model's feature preservation ability. A model fingerprint feature extraction network (CNN) is constructed as follows:

[0140] The fingerprint feature extraction network of the model is a deep convolutional neural network, with the following structure:

[0141] Input layer: Image size normalized to 224×224, RGB three channels;

[0142] Convolutional Module (ConvModule):

[0143] Conv1: Convolutional kernel 3×33, output channels 64, activation function ReLU;

[0144] Conv2: Convolution kernel 3×33, output channels 128, activation function ReLU;

[0145] Conv3: Convolutional kernel 3×33, output channels 256, activation function ReLU;

[0146] MaxPooling module: Performs 2×22 max pooling after every two convolutional layers.

[0147] Fully Connected Module (FCLayer):

[0148] Flatten layer;

[0149] FC1: Output dimension is 1024, activation function is ReLU;

[0150] Dropout: Prevents overfitting, with a retention rate of 0.5;

[0151] Feature Output Layer: The FC1 layer is taken as the intermediate layer output and used as the fingerprint feature vector of the frame-level model;

[0152] The network training data consists of multiple real AI-generated outfit video samples and non-AI video samples. The training objective is to maximize the distinguishability of the generated models. The loss function is TripletLoss, which is optimized by inputting positive and negative sample triples.

[0153] The generation of the current video model fingerprint vector VA:

[0154] The keyframe set F of the video to be verified is input into the above convolutional neural network (CNN) to obtain the intermediate layer output feature vector vi∈R for each frame. 1024 Perform feature aggregation processing on the feature vectors of all n=10 frames: ;

[0155] Then, L2 normalization is performed on the aggregated feature vector Vraw to obtain the model fingerprint vector in unit vector form:

[0156] VA = Vraw / ||Vraw|| 2 Where ||Vraw|| represents the magnitude of the aggregated feature vector;

[0157] The VA∈R 1024 This is the model fingerprint vector of the current outfit video;

[0158] Generation and writing of fingerprint vector VB in the registration model:

[0159] During the outfit video generation stage, the generation end performs the same processing flow as described above (keyframe extraction, CNN network, aggregation, normalization) on the generated video content to obtain the generation end registered model fingerprint vector VB. The generation network structure, parameters, keyframe extraction rules and feature aggregation method of the registered model fingerprint vector VB are consistent with those of the model fingerprint vector VA to ensure subsequent comparability.

[0160] Subsequently, the system will write the registration model fingerprint vector VB along with the following data into the registration traceability record set:

[0161] Video unique identifier ID, video generation model ID, watermark reference sequence WB, video generation timestamp, and generation parameter summary (such as clothing tags and environment settings).

[0162] When deploying a blockchain module, the traceability record containing VB can be hashed and written onto the chain to enhance the immutability and legal validity of the registered data.

[0163] S4 also includes S42;

[0164] S42. Based on the model fingerprint vector VA corresponding to the obtained video to be verified and the registered model fingerprint vector VB, perform cosine similarity calculation, output similarity value SIM, and quantitatively analyze the similarity between the AI-generated outfit video and the level model.

[0165] The similarity value (SIM) is calculated and output using the following algorithm formula;

[0166] ;

[0167] In the formula, SIM(VA, VB) represents the similarity value between the model fingerprint vector VA and the registered model fingerprint vector VB, and ||VA|| and ||VB|| represent the magnitudes of the model fingerprint vector VA and the registered model fingerprint vector VB, respectively;

[0168] The model fingerprint vector similarity calculation formula SIM used in this formula is based on the classic cosine similarity formula in mathematics. This formula is widely used to calculate the cosine value of the angle between two vectors in vector space to measure directional similarity rather than magnitude difference.

[0169] Why use cosine similarity? Cosine similarity is suitable for evaluating the directional consistency between two high-dimensional vectors, eliminating the influence of amplitude.

[0170] In model fingerprinting applications, the model output vector has been L2 normalized, and cosine similarity can accurately assess the consistency of the "generative model style".

[0171] The practical business implications of this formula:

[0172] If two videos exhibit a high degree of consistency in directionality in their keyframe feature responses, it indicates that they come from the same generative model.

[0173] Conversely, if the similarity is too low, even if the appearance style is similar, it can be identified as fake or generated by swapping sources.

[0174] In this embodiment, a model fingerprint feature extraction network based on a convolutional neural network (CNN) is introduced to perform deep feature extraction and aggregation on key frames in AI-generated outfit videos, generating a structural model fingerprint vector (VA), thereby achieving a unique identifier for the generation model level. This processing method can effectively extract the intrinsic modeling features of the "generation source" even when the video content surface is highly similar, overcoming the problem that traditional video similarity assessment cannot distinguish between "same-origin forgeries". During the generation of the model fingerprint vector VA, standard sampling rules are used to extract frame images with the motion features of the outfit subject from the key frame set F, which can significantly enhance the sensitivity of the feature expression to "structural differences in the generation model". For example, different AI models have potential differences in their expression when processing clothing texture and fabric motion blur. These details can be captured by the high-dimensional response of the CNN intermediate layer, thus forming a highly discriminative fingerprint vector. The registration process of the model fingerprint vector VB is completely consistent with that of VA, ensuring the comparability of subsequent SIM similarity calculations. Using classical cosine similarity for vector similarity analysis can effectively eliminate amplitude interference, focusing only on feature direction, ensuring that "whether it is generated from the same source" can still be accurately identified under different generation conditions. This setting avoids the problem of overall amplitude differences caused by image compression or post-processing. The essential effect of this strategy is that even in the face of tampering such as forgery, model swapping, and editing, verification cannot be passed at the model fingerprint level if the original generation model is not used. Physically speaking, the output of the similarity value SIM essentially establishes a mapping path from the image data space to the model architecture space, forming a "deep source tracing trajectory." This achieves a verification leap from content consistency to model consistency, greatly improving the credibility generation capability and forgery defense strength of AI outfit videos.

[0175] Example 6

[0176] Please see Figure 1 Specifically: S5 includes S51;

[0177] S51. Using a pre-built video generation model fingerprint sample library, statistically analyze the similarity value SIM interval of a large number of registered AI-generated outfit videos. Based on the lower bound of the interval where the statistical distribution is at least 95% confidence level, set the similarity threshold Sth for the model fingerprint. Then, evaluate the watermark similarity between the real-time acquired model fingerprint vector VA and the registered model fingerprint vector VB and the similarity threshold Sth.

[0178] Meanwhile, the generated model fingerprint vector VA, the registered watermark reference sequence WB, and the generation timestamp extracted from the AI-generated outfit video to be verified are verified on-chain. On-chain verification is carried out by constructing a traceability hash value and submitting it to the on-chain database node. The database node then performs parallel verification on the hash matching, the validity of the registered generator signature, and the logical order of the timestamp. If all three verifications pass, the on-chain verification is reported as passed; if any one verification fails, the on-chain verification is reported as failed.

[0179] The specific assessment content is as follows:

[0180] When the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is greater than or equal to the similarity threshold Sth, and the verification is passed, it is determined that the two are generated from the same source, and the current AI-generated outfit video is marked as having passed the similarity verification.

[0181] If the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is less than the similarity threshold Sth, and the verification is passed, it is determined that there is tampering, and the current AI-generated outfit video is marked as a similarity verification failure.

[0182] When on-chain verification fails, the registration is deemed unverifiable, and the current AI-generated outfit video is marked as having an unknown source.

[0183] S5 also includes S52;

[0184] S52. Based on the judgment results of watermark similarity assessment and on-chain verification, execute the control actions corresponding to the verification status, specifically including:

[0185] When the similarity verification of the judgment result is passed, the corresponding AI-generated outfit video will be marked as publicly shared and the video platform will be allowed to open the dissemination permission;

[0186] When the similarity verification fails, the content source tracing manual review process is triggered.

[0187] When the on-chain verification result is unsuccessful, the corresponding AI-generated outfit video will be directly identified as a video of unknown origin, and risk labeling will be applied to the AI-generated outfit video to restrict its dissemination and use, triggering a manual review process for content tracing.

[0188] In this embodiment, the method employs a statistical approach to construct a model fingerprint similarity threshold Sth. This threshold is set based on a large number of registered samples within a 95% confidence interval, effectively avoiding misjudgments caused by fluctuations in model features and improving the accuracy of fingerprint matching. For example, if an empirical setting method is used, a specific case might lead to the wrong labeling of a legitimate video as abnormal. However, this statistical threshold setting method ensures a high recognition success rate for the vast majority of genuinely generated content.

[0189] The on-chain verification mechanism introduces a joint hash verification strategy using three elements: generation timestamp, model fingerprint vector (VA), and registration watermark reference sequence (WB). This enhances the non-repudiation of registration information in both spatial location (data content) and temporal dimension (generation order). For example, if video content is copied and tampered with locally, even if the appearance is similar, it cannot be matched with the original registration value through on-chain hash value, effectively blocking offline counterfeit verification attempts. Ultimately, the dual judgment results of similarity assessment and on-chain verification are used to control action triggering, with a clear layered security strategy: if it passes completely, it can be shared openly, ensuring the normal dissemination of trustworthy content; if only the similarity does not match, it enters the review process to avoid mistakenly penalizing legitimate content; if on-chain verification fails, it is immediately marked as having an unknown source, blocking its further dissemination path. This strategy forms a content credibility process from "assessment, verification, and control," ensuring efficient processing and improving the ability to identify and handle illegal counterfeiting, model translation, and tampered videos.

[0190] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal source tracing method for AI-generated outfit videos, characterized in that: Includes the following steps: S1. Deploy multimodal acquisition nodes on the video generation chain to collect the original video dataset of AI-generated outfit videos in real time, and upload the original video dataset to the edge tracing node; perform normalization processing on the original video dataset in the edge tracing node to obtain a normalized outfit video dataset; S2. Based on the standardized dressing video dataset, generate a registered watermark reference sequence WB and register the registered watermark reference sequence WB in the traceability record database; S3. Extract the watermark bit sequence WA based on the video data to be verified, and then calculate the error rate with the registered watermark reference sequence WB to obtain the watermark bit error rate BER. Then compare the watermark bit error rate BER with the preset watermark interval threshold to determine the current watermark verification status and trigger fingerprint depth verification processing. S4. When fingerprint depth verification is triggered, perform fingerprint depth verification, extract the model fingerprint vector VA corresponding to the video to be verified, and calculate the similarity value SIM with the registered model fingerprint vector VB. S5. Then, preset the similarity threshold Sth and the similarity value SIM to evaluate the watermark similarity, and execute the corresponding control actions based on the evaluation results.

2. The multimodal source tracing method for AI-generated outfit videos according to claim 1, characterized in that: S1 includes S11; S11. In the outfit video generation system, a multimodal acquisition node is set between the generation output end and the encoding output end on the AI-generated outfit video generation link. Through the multimodal acquisition node, multimodal data is collected during the generation and verification process of AI-generated outfit videos to obtain the original video dataset. The original video dataset is then uploaded to the edge tracing node through the edge communication interface. The multimodal acquisition nodes include video image acquisition points, audio data acquisition points, and text data acquisition points; The original video dataset includes a set of joint frames F, an optical flow field set O, a saliency map set M, a DCT coefficient set C, an audio spectrogram set A, and a subtitle timeline set T.

3. The multimodal source tracing method for AI-generated outfit videos according to claim 1, characterized in that: S1 further includes S12; S12. In the edge tracing node, the original video dataset is normalized to obtain a standardized outfit video dataset. The normalization process includes video normalization, audio normalization, and subtitle text normalization. The video normalization process scales the video frames to a preset standard size of 4K using data collected from video capture points, then resamples the video frame rate to a uniform frame rate of 60fps, and finally converts the original image data into a standard color space of RGBBT.

709. The audio normalization process involves performing sampling rate unification processing to unify the audio sampling rate to a standard value of 44100Hz, and then performing channel format normalization processing to downmix multi-channel audio into mono or stereo format. The subtitle text normalization process extracts the start and end time information of the subtitles to form a subtitle timeline set T, and calibrates the subtitle display nodes based on the video frame timeline and the audio timeline.

4. The multimodal source tracing method for AI-generated outfit videos according to claim 3, characterized in that: S2 includes S21; S21. Based on the standardized outfit video dataset, watermark embedding processing is performed on the main clothing area and dynamic area in the outfit video. The watermark embedding processing is performed by executing a multimodal watermark embedding strategy set. Based on the execution result of the multimodal watermark embedding strategy set, the embedded watermark bit sequence is extracted from the standard frame position and frequency domain position during the embedding process, a registration watermark reference sequence WB is generated, and a registration operation is performed. The multimodal watermarking embedding strategy set includes region selection strategy K1, time-domain diffusion strategy K2, frequency-domain embedding strategy K3, dynamic adaptive strategy K4, and error correction coding strategy K5; The registration operation combines the registration watermark reference sequence WB with the corresponding outfit video's generation model identifier, generation parameter summary, generation timestamp, and generation metadata to form a traceability data summary; The trusted execution environment at the generation end performs digital signatures on the registered watermark reference sequence WB and the source data digest; The signed registration watermark reference sequence WB and the generated metadata are written into the registration traceability record set as registration traceability records and simultaneously written into the blockchain ledger.

5. A multimodal source tracing method for AI-generated outfit videos according to claim 4, characterized in that: S3 includes S31; S31. Based on the AI-generated outfit video to be verified, extract the watermark extraction sequence WA using the same multimodal acquisition node; then calculate the watermark bit error rate BER of the current AI-generated outfit video by comparing the watermark extraction sequence WA with the registered watermark reference sequence WB.

6. The multimodal source tracing method for AI-generated outfit videos according to claim 4, characterized in that: S3 further includes S32; S32. In the edge tracing node, two watermark interval thresholds related to watermark verification are pre-set. The watermark interval thresholds include the watermark passing threshold Bth1 and the strong verification trigger threshold Bth2. Then, the real-time watermark bit error rate BER is compared with the two watermark interval thresholds to determine the current watermark verification status and trigger model fingerprint analysis and blockchain trust verification. The specific comparison content is as follows: When the watermark bit error rate BER ≤ watermark passing threshold Bth1, the watermark consistency of the current AI-generated outfit video to be verified is determined to meet the passing condition, the watermark passing status result is output, and the AI-generated outfit video to be verified is marked as a watermark verification passed video. When the watermark passes the threshold Bth1 < watermark bit error rate BER ≤ strong verification trigger threshold Bth2, it is determined that the current AI-generated outfit video to be verified is in an uncertain state of watermark verification, the watermark gray area state result is output, and fingerprint depth verification processing is triggered. When the watermark bit error rate BER is greater than the strong verification trigger threshold Bth2, the watermark consistency of the current AI-generated outfit video to be verified is deemed insufficient, the watermark verification failure result is output, and fingerprint deep verification processing is forcibly triggered.

7. A multimodal source tracing method for AI-generated outfit videos according to claim 6, characterized in that: S4 includes S41; S41. When fingerprint depth verification is triggered, extract the model fingerprint vector VA of the outfit video generated by the AI ​​to be verified and the registered model fingerprint vector VB; the specific extraction content is as follows; Based on the standardized outfit video dataset corresponding to AI-generated outfit videos, key frames representing the outfit scene content are extracted from the key frame set F; Keyframes representing the content of the outfit image are input into a pre-defined model fingerprint feature extraction network to extract the intermediate layer feature responses of the model fingerprint feature extraction network. Feature aggregation processing is performed on the intermediate layer feature responses corresponding to multiple representative keyframes, and the aggregation results are normalized and dimensionally constrained to generate the fingerprint vector VA of the current video model. In the outfit video generation stage, based on the same network structure, key frame selection rules and feature processing flow as the model fingerprint feature extraction network, the generated outfit video is subjected to model fingerprint extraction processing to generate the registered model fingerprint vector VB. The registration model fingerprint vector VB, along with the corresponding outfit video identifier information, watermark reference sequence, and generated metadata, are written into the registration traceability record set. At the same time, the unique video identifier ID of the outfit video to be verified is read from the registration and traceability record set, and the registration model fingerprint vector VB, which was registered synchronously when the video was generated, is read through the unique video identifier ID. The registration model fingerprint vector VB is a vector generated by the generator based on the same features extracted by the network and rules as the model fingerprint vector VA, and is encrypted, signed and stored in the central database or blockchain node.

8. A multimodal source tracing method for AI-generated outfit videos according to claim 7, characterized in that: S4 also includes S42; S42. Based on the model fingerprint vector VA corresponding to the obtained video to be verified and the registered model fingerprint vector VB, perform cosine similarity calculation, output similarity value SIM, and quantitatively analyze the similarity between the AI-generated outfit video and the level model.

9. A multimodal source tracing method for AI-generated outfit videos according to claim 8, characterized in that: S5 includes S51; S51. Using a pre-built video generation model fingerprint sample library, statistically analyze the similarity value SIM interval of a large number of registered AI-generated outfit videos. Based on the lower bound of the interval where the statistical distribution is at least 95% confidence level, set the similarity threshold Sth for the model fingerprint. Then, evaluate the watermark similarity between the real-time acquired model fingerprint vector VA and the registered model fingerprint vector VB and the similarity threshold Sth. Simultaneously, the generated model fingerprint vector VA, the registered watermark reference sequence WB, and the generation timestamp extracted from the AI-generated outfit video to be verified are performed on-chain. The on-chain verification is achieved by constructing a traceability hash value and submitting it to the on-chain database node. The database node then performs parallel verification on the hash matching, the validity of the registered generator signature, and the logical order of the timestamp. If all three verifications pass, the on-chain verification is reported as passed; if any one verification fails, the on-chain verification is reported as failed. The specific assessment content is as follows: When the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is greater than or equal to the similarity threshold Sth, and the verification is passed, it is determined that the two are generated from the same source, and the current AI-generated outfit video is marked as having passed the similarity verification. If the similarity value SIM(VA, VB) between the model fingerprint vector VA and the registered model fingerprint vector VB is less than the similarity threshold Sth, and the verification is passed, it is determined that there is tampering, and the current AI-generated outfit video is marked as a similarity verification failure. When on-chain verification fails, the registration is deemed unverifiable, and the current AI-generated outfit video is marked as having an unknown source.

10. A multimodal source tracing method for AI-generated outfit videos according to claim 9, characterized in that: S5 also includes S52; S52. Based on the judgment results of watermark similarity assessment and on-chain verification, execute the control actions corresponding to the verification status, specifically including: When the similarity verification of the judgment result is passed, the corresponding AI-generated outfit video will be marked as publicly shared and the video platform will be allowed to open the dissemination permission; When the similarity verification fails, the content source tracing manual review process is triggered. When the on-chain verification result is unsuccessful, the corresponding AI-generated outfit video will be directly identified as a video of unknown origin, and risk labeling will be applied to the AI-generated outfit video to restrict its dissemination and use, triggering a manual review process for content tracing.

Images