Real face digital human multi-layer superposition fingerprint authorization and right protection management method and system and storage medium

Abstract

This invention discloses a method, system, and storage medium for multi-layered fingerprint authorization and rights protection management of real-person facial digital humans. The method includes: completing static facial images, liveness recordings, and initial modeling within the same session and generating a data acquisition and binding package; submitting a disclosure package and completing informed consent through keyword highlighting interaction in a controlled environment; distributing the data to open distribution nodes after signing by the issuing node to form an authorization status matrix; querying the authorization status and obtaining a short-term token before use by the user; fine-tuning the model and then transmitting it back and overlaying a new fingerprint layer; detecting unauthorized use from the content and model layers, and reconstructing the lineage path by parsing the multi-layered fingerprint chain layer by layer in the model layer; and automatically initiating rights protection after securing evidence in a trusted environment and pushing it to a third-party cloud notary office to obtain a timestamp and notarized certificate. This invention achieves closed-loop management of real-person facial digital humans from data acquisition, authorization, use, withdrawal to infringement detection and rights protection evidence collection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence digital human technology, and in particular to a method, system and storage medium for multi-layer overlay fingerprint authorization and rights protection management of real-face digital humans, covering the entire lifecycle management of real-face data collection and archiving, informed consent storage, cross-platform authorization status distribution, fine-tuning and transmission of lineage tracking, infringement detection and rights protection evidence collection. Background Technology

[0002] With the development of generative artificial intelligence technology, digital humans based on real human faces are increasingly being used in film and television production, advertising endorsements, and virtual live streaming. However, the current use of digital human faces faces the following problems: First, there is a lack of credible authorization evidence during the data collection phase. Existing liveness detection technologies are mainly geared towards identity verification scenarios. They do not generate commercially viable digital asset models during data collection, nor do they cryptographically link the collection behavior with authorization intent. This makes it difficult to prove afterward whether the target individual was actually present and consented to the authorization.

[0003] Second, there is a lack of cross-platform synchronization mechanisms for authorization status. Existing digital human authorizations mostly rely on paper contracts or point-to-point authorizations. When multiple platforms use the same digital human, changes in authorization status cannot be synchronized to each platform in real time, resulting in the continued use of digital humans that have been revoked.

[0004] Third, the informed consent process lacks technical evidence. Whether the rights holder actually read the script and fully understood the role when granting authorization currently relies solely on contract terms, lacking technical records of the reading process, making it difficult to provide evidence in case of disputes later.

[0005] Fourth, infringement detection methods are limited and cannot track secondary development. Existing technologies mostly focus on overall facial similarity comparison, lacking effective detection methods for partial feature theft and model distillation theft; when the thief modifies and creates a new model based on the original model, existing watermarking schemes become ineffective after fine-tuning or distillation.

[0006] Fifth, there is a lack of technical support for rights protection and evidence collection. After infringement is discovered, the processes of evidence collection, preservation, platform notification, and compensation calculation all rely on manual operation, which is inefficient and results in insufficient evidentiary value. Summary of the Invention

[0007] The technical problem this invention aims to solve is: how to provide a management method and system that covers the entire lifecycle of digital human faces, from collection and authorization to infringement protection, so that the authorization process is traceable, the authorization status can be synchronized across platforms in real time, withdrawal can be technically executed, secondary development models can trace lineage through multi-layered fingerprints, and infringement can be automatically detected and form a judicially recognized chain of evidence.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for multi-layer overlay fingerprint authorization and rights protection management of a real-person digital human, including seven steps: data collection and filing, informed consent, filing and distribution, authorized access, fine-tuning and feedback, infringement detection and rights protection execution, the latter steps depend on the data produced by the aforementioned steps.

[0009] Beneficial effects: 1. The data collection and binding package cryptographically binds the modeling materials, the liveness intent, and the initial model, solving the problem of proving "who authorized it"; 2. The multi-layer fingerprinting mechanism adds a layer to the model with each authorization and fine-tuning, ensuring that the model's lineage follows wherever it flows, thus solving the problem of tracking theft during secondary development; 3. Keyword highlighting and interactive review provide sentence-by-sentence interactive evidence of the informed consent process, resolving the controversy over "whether the informed consent was fully obtained"; 4. The distributed authorization state network enables real-time cross-platform synchronization of authorization changes, solving the difficulty of withdrawing execution; 5. The entire process from evidence collection to notification is automated, and the use of timestamps from third-party cloud notary offices enhances the validity of the evidence. Detailed Implementation

[0010] Example 1: System Architecture. This system includes a data collection and filing module, an informed consent module, a filing and distribution module, an authorization and invocation module, a fine-tuning and feedback module, an infringement detection module, and a rights protection execution module. The data collection and filing module is deployed on the data collection end, the informed consent module is deployed on the review end, the filing and distribution module adopts a distributed architecture, the authorization and invocation module is provided to various application platforms in the form of APIs, the infringement detection module connects to content crawlers and model detection tools, and the rights protection execution module integrates a trusted browser, a judicial link interface, a cloud notary office interface, and a platform complaint interface. The overall system architecture is as follows: Figure 1 As shown.

[0011] Example 2: Data Acquisition and Archiving. At the acquisition end, the target individual is guided to capture multi-angle static facial images under specified lighting and distance conditions using multiple cameras. Simultaneously, they are guided to perform preset actions and read preset statements bound to the intended authorization scope. Liveness recordings are made throughout the process, and voiceprint features are collected. An initial 3D facial model is generated in real-time based on the static images. SHA-256 digests are calculated for the three types of products, and after concatenation, a hash is calculated again to generate a joint binding fingerprint. This fingerprint, along with the timestamp, device fingerprint, and sensor parameters, forms the acquisition binding package, such as... Figure 2 As shown. The revocation trigger key is generated by the rights holder through facial recognition and stored separately from the binding package.

[0012] Example 3: Informed Consent and Sealed Review. The user submits a disclosure package containing script hashes and character positioning. The rights holder logs into the review environment via a trusted terminal. The script is encrypted and stored on the server, and during review, it is decrypted and rendered in memory segment by segment. Screenshots and copying are disabled in the environment, and a dynamic watermark containing the reviewer's ID and timestamp is overlaid on the display layer. The system classifies content according to sensitivity. After the script text is loaded, the system randomly extracts keywords and phrases from each line as highlight marks. The highlights are embedded in the text lines in a fixed style, without suddenly popping up or obscuring the main text. When the user clicks on a highlight mark during normal reading, the reading progress unlocks and advances with each clicked highlight. Content before a highlight cannot be scrolled to the next screen. The confirmation button is activated only after all highlight marks have been clicked and the user has read to the end; key plot sections and sensitive content sections additionally require a complete signed confirmation mark, such as... Figure 3 As shown. All line-by-line interaction traces are recorded and hashed for evidence. Upon review, a signed informed consent form is obtained through both facial recognition and voiceprint verification. Version differences are automatically generated when the script is updated; only the changed portions are reviewed.

[0013] Example 4: Distributed Authorization State Network. Issuing nodes undergo qualification verification, while distribution nodes can run arbitrarily. Before making a call, the user submits a facial identifier and usage scenario to the distribution node, obtaining a self-contained signed authorization credential and a short-term token (valid for 15 minutes). The user verifies the signature using their local root public key and does not trust the distribution node. State changes are notified through a three-tiered system: Webhook push, incremental polling, and public logs. Figure 4 As shown.

[0014] Example 5: Using Control and Fine-tuning Backhaul. Models are stored hierarchically according to authorization levels, and the corresponding level is dynamically loaded during rendering. Model inference does not extend beyond the domain; identity watermarks are embedded only when local deployment is required. Fine-tuning is divided into five levels (M0 read-only to M4 deep modification), with each level limiting the extent of modification. After fine-tuning, the model must be backhauled to calculate similarity with the parent model, overlay a new fingerprint layer, establish a lineage tree, and incorporate the derived model into the detection library.

[0015] Example 6: Multi-layered fingerprint overlay. The first layer of archived fingerprint is generated and bound to the rights holder during collection; a second layer of authorized fingerprint is overlaid each time authorization is granted, recording the publisher, user, scenario, and level; after each minor adjustment, a new layer of secondary development fingerprint is overlaid, recording the developer, modification level, and similarity; all overlay layers are chained together, with each subsequent layer referencing the summary of the previous layer, and all are returned to the publisher's archive. During infringement detection, the fingerprint chain is analyzed layer by layer to reconstruct the complete lineage path; a broken chain or a mismatch with the archive indicates misappropriation. Figure 5 As shown. Meanwhile, the model output retains statistical perturbations as a fallback, supplementing detection with probe behavioral fingerprints and output distribution distances.

[0016] Example 7: Authorization State Machine. The authorization state includes five states: pending review, authorized, in use, revoked, and expired. State transitions are as follows: Figure 6 As shown. When a user invokes the digital human, the user must be in an "authorized" or "in use" state. After withdrawal, the state changes to "withdrawn," and all invocations are rejected.

[0017] Example 8: Withdrawal Notification and Execution. After the rights holder initiates withdrawal, the system changes the authorization status to "Withdrawn" and broadcasts the withdrawal notification to all distribution nodes through the distributed status network. The withdrawal notification includes a timestamp, scope identifier, and effective time. Application platforms with configured Webhook receive the notification in real time and return a confirmation receipt; platforms without configured Webhook automatically reject the request upon finding the "Withdrawn" status on the next call. Platforms that do not return a receipt are continuously retried and recorded by the system. After confirmation by each platform, execution proceeds in four layers: stop new generation, destroy materials, withdraw unreleased content, and withdraw released content. Each work is verified to ensure it falls within the scope of informed consent; within the scope, the contract is locked and the content is not removed; outside the scope, a 15-working-day removal window period begins. The withdrawal notification, platform receipts, and execution results are all written to the status change log and a timestamp is simultaneously applied for for evidence storage. Figure 7 As shown.

[0018] Example 9: Infringement Detection. Content Layer: After cross-platform crawling, sub-feature comparison and multi-dimensional identifiability assessment are performed. Sub-features include at least eyes, nose, mouth, contour, voiceprint, and body posture. Model Layer: Multi-layer superimposed fingerprint chain analysis is used. The bottom layer supplements detection through probe behavioral fingerprints and output distribution distance calculation. Human Identification Layer: The original static photos and live video frames of the rights holder are extracted from the collection and binding package and displayed side-by-side with the faces in the suspicious content for reviewers. Third-party observers unfamiliar with the rights holder perform visual comparison to assess whether ordinary people can identify the natural person, forming a human identification conclusion and recording it. Import Sampling Comparison: Rights holders or rights defenders can upload suspicious images, videos, or model files through the platform entry. The system automatically extracts facial feature vectors, sub-features, and model fingerprints from them, and performs side-by-side comparison and analysis with the original files and multi-layer superimposed fingerprint chains in the collection and binding package. It outputs similarity scores, sub-feature difference heatmaps, and fingerprint chain lineage paths, and supports manual annotation and conclusion archiving. The comprehensive judgment process is as follows: Figure 8 As shown.

[0019] Example 10: Rights Protection Enforcement and Cloud Notary Access. A trusted virtual browser fully crawls infringing content, calculates its hash, and pushes it to the judicial chain and cloud notary institutions, applying for an RFC 3161 timestamp. Infringement notices are automatically generated and submitted through the platform interface, with responses tracked. Counter-notifications are automatically verified. Losses are calculated based on duration and reach, such as... Figure 9 As shown, the cloud-based notary office has no equity affiliation with the platform operator, and electronic notarized documents can be applied for with one click during the litigation stage.

[0020] The system provides four types of access ports: application platform port (six types of platforms, including short dramas, advertisements, games, live streaming, e-commerce, and social media, each with its corresponding API), regulatory department port (filing audit, violation alerts, and forced removal), legal department port (evidence package retrieval, evidence storage and verification, and dispute management), and cloud notary office port (RFC 3161 timestamp application, electronic notarized document issuance, and evidence storage and verification).

[0021] Figure 1 This is a system architecture diagram of the multi-layer overlay fingerprint authorization and rights protection management method for real-person digital human faces provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the data acquisition and binding package generation process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the sealed review environment and keyword highlighting interaction confirmation provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the distributed authorization state network topology provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a multi-layered fingerprint chain structure provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the state transition of the authorization state machine provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the withdrawal notification and execution process provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the multi-source infringement detection process provided in the embodiments of the present invention, which includes a content layer, a model layer, a bottom layer, a manual identification layer, and an import sampling comparison layer; Figure 9 This is a schematic diagram of the rights protection execution and cloud notarization access process provided in the embodiments of the present invention.

Claims

1. A method, system, and storage medium management method for multi-layer overlay fingerprint authorization and rights protection management of real-person facial digital humans, characterized in that: include: S1 Data Acquisition and Archiving Steps: In the same acquisition session, continuously complete the acquisition of multi-view static facial images of the target natural person, the acquisition of live video recordings including dynamic expressions and voiceprint features, and the on-site initial 3D facial modeling based on the static facial images; calculate cryptographic digests for the static facial images, live video recordings, and initial 3D facial models respectively, generate a joint binding fingerprint, and form an acquisition binding package together with the acquisition timestamp, device identifier, and authorization intention record confirmed by the live person and persistently store it; S2 Informed Consent Steps: The user submits a structured disclosure package to the rights holder, which includes a script content hash, character positioning description, usage scenarios, online platform scope, and project duration; the user displays the disclosed content to the rights holder in a controlled review environment, records the rights holder's reading behavior of each content paragraph, and requires confirmation of each paragraph marked as key plot and sensitive content; after liveness verification, an informed consent certificate is generated, which includes an authorization clause that the rights holder will not request the removal of the already published work if the use does not exceed the disclosure scope; S3 Filing and Distribution Steps: After the collection binding package and informed consent certificate are signed by the issuing node, they are broadcast to the distribution node through the incremental synchronization protocol to form an authorization status matrix that can be queried in real time; S4 Authorization Invocation Steps: Before invoking the digital human face, the user submits the face identifier, usage scenario, and caller identifier to the distribution node to obtain an authorization status certificate and a short-term access token signed by the issuing node; the user verifies the signature chain of the authorization status certificate locally and binds the query log to this generated invocation for record keeping; S5 Fine-tuning and Feedback Steps: After fine-tuning the digital human face model within the authorized scope, the user sends the fine-tuned model or its feature summary back to the filing platform; the platform calculates the similarity between the fine-tuned model and the parent model, verifies whether the modification range is within the authorized scope, overlays a new model fingerprint layer, and establishes a derivative relationship; S6 Infringement Detection Steps: Multi-source detection of unauthorized use, including at least one of the following: (a) Comparison of facial sub-feature vectors and multi-dimensional identifiability assessment of images and videos collected by the content platform; (b) Multi-layer superimposed fingerprint chain analysis, passive fingerprint detection, probe behavioral fingerprint comparison, and output distribution distance calculation for suspicious models; (c) Extracting the original static profile and liveness video frames of the rights holder from the S1 collection and binding package, displaying them side-by-side with the faces in the suspicious content, and having the reviewer visually compare them to assess natural recognizability—that is, whether an ordinary observer unfamiliar with the rights holder can identify the face in the suspicious content as that natural person, forming a manual identification conclusion and recording it; (d) The system provides an entry point for importing sampled works for comparison. After rights holders or rights defenders upload suspicious images, videos, or model files, the system automatically extracts facial feature vectors, sub-features, and model fingerprints. These are then compared side-by-side with the original files and multi-layered fingerprint chains in the data collection and binding package. The system outputs a similarity score, a sub-feature difference heatmap, and a fingerprint chain lineage path, and supports manual annotation and conclusion archiving. The rights holder comparison benchmark used for infringement detection comes from the data collection and binding package of S1. S7 Rights Protection Execution Steps: When suspected infringement is detected, the infringing content is completely captured in a trusted environment and its cryptographic hash is calculated. The data is then written into a judicially recognized blockchain to generate a certificate of evidence storage. An infringement notice is automatically generated based on the evidence storage results, submitted through the platform interface, and the response is tracked. The validity of the authorization certificate is automatically verified in response to the counter-notification from the accused party. The scope of authorization used in the S6 infringement determination is derived from the informed consent certificate in S2, and the authorization status queried in the counter-notification verification is derived from the authorization status matrix in S3.

2. The method according to claim 1, characterized in that, In S1: the live video recording includes preset actions and preset statements to be read aloud to guide the target natural person, and the content of the preset statements is bound to the scope of authorization; the collection environment parameters include lighting conditions, collection distance and sensor model, which are written into the collection binding package for anti-counterfeiting; the same natural person generates independent collection binding packages at different age groups to form an appearance timeline.

3. The method according to claim 1, characterized in that, In S2: the disclosed content is classified into regular plot segments, key plot segments, and sensitive content segments according to sensitivity; the controlled review environment prohibits the export, screenshotting, and copying of content, and the display layer is overlaid with a dynamic watermark containing the reviewer's identifier and timestamp; after the script text is loaded, the system randomly extracts keywords and phrases from each line as highlight marks, and the highlight marks are embedded in the text lines in a fixed style, without suddenly popping up, obscuring the main text, or interrupting the reading flow; When a user clicks a highlighted mark during normal reading, the reading progress will unlock and advance as the highlighted mark is clicked. Content before a highlighted mark that has not been clicked cannot be scrolled to the next screen. The confirmation button is activated only after all highlighted marks have been clicked and the reader has scrolled to the end of the text; each click of a highlighted mark and its corresponding timestamp are recorded and hashed for evidence, forming a sentence-by-sentence interaction trajectory; for key plot segments and sensitive content segments, in addition to clicking each sentence to highlight them, the rights holder is required to sign and confirm the entire segment; when the script is updated, version differences are automatically generated and only the changed parts are required to be reviewed.

4. The method according to claim 1, characterized in that, In S3: the issuing node needs to be qualified, and the distribution node can be run by any entity; the authorization status credential is a self-contained signature structure, and the user completes the verification through the locally preset root signature public key, without relying on the trustworthiness of the distribution node; the authorization status matrix is ​​maintained independently according to the scenario dimension; the validity period of the short-term access token does not exceed 15 minutes; status changes are notified through a three-layer mechanism of real-time push, incremental polling and publicly verifiable status logs.

5. The method according to claim 1, characterized in that, In S5: fine-tuning authorization is divided into read-only file, expression-driven file, appearance replacement file, facial fine-tuning file and deep modification file, each file corresponds to different model layer open permissions and modification range thresholds; After fine-tuning, the system automatically calculates the facial similarity to the parent model and triggers an alert if the similarity falls below the authorized threshold. The derived models are incorporated into the infringement detection and comparison library, forming a lineage tree from the collected and bound root model to the derived models at all levels.

6. The method according to claim 1, characterized in that, The S6 employs a multi-layered fingerprint mechanism for lineage tracing. The first layer is the archived fingerprint. When collecting and archiving data in S1, the first fingerprint is calculated on the initial 3D facial model and bound to the rights holder's identity information as the fingerprint chain root. The second layer is the authorization fingerprint. When S4 authorization is invoked, the second fingerprint is superimposed on the model delivered to the application party on the basis of the first fingerprint. It includes the digital publisher identifier, application party identifier, authorization scenario, timestamp and authorization level. After superposition, it is sent back to the digital publisher archive. The third and subsequent layers are secondary development fingerprints. When the S5 fine-tuning is returned, the application adds a new layer on the previous fingerprint after each fine-tuning. This includes the developer identifier, modification level, similarity after modification, and fine-tuning timestamp. After the layering, the digital publisher archive is returned. Each development or authorization is completed by adding a layer, and N developments result in N layers being added. The layers are linked together in chronological order, with each layer referencing the fingerprint digest of the previous layer. During infringement detection, the fingerprint chain of the model to be detected is analyzed layer by layer, and the complete lineage path is reconstructed by tracing back from the latest layer. If the fingerprint chain is broken, a layer is missing, or a record of a certain layer does not match the digital publisher's archive, it is determined that the model is being used without authorization. The multi-layered fingerprint is embedded in the model weight layer in an invisible manner, without affecting the inference effect, and the new layer does not cover the existing layer; at the same time, the statistical perturbation mark is retained in the model output as a fallback, and supplementary detection is performed by calculating the probe behavior fingerprint and the output distribution distance.

7. The method according to claim 1, characterized in that, In S7: the trusted environment is a notarized virtual browser environment that fully captures infringing pages, original media files, publisher information, and interactive data, and records the entire capture process; the infringement notice includes rights holder information, infringement comparison analysis, infringing content location, and legal basis; the counter-notification verification includes verification of the validity of the authorization certificate signature, the coverage of the usage scenario, and the authorization time; and the infringement loss is automatically calculated based on the duration and scale of the infringement.

8. The method according to claim 1, characterized in that, In S7: the evidence consolidation results are pushed to an independent third-party cloud notary office through an encrypted channel; during evidence consolidation, an RFC 3161 standard timestamp is applied for from the cloud notary office, and the timestamp is bound to the evidence package hash; timestamps are applied for simultaneously during data collection, filing, and informed consent signing; during the rights protection and litigation stage, multiple evidence consolidation records, timestamps, and authorization change records are summarized and an electronic notarized certificate is issued through an interface application; the cloud notary office has no equity relationship with the platform operator.

9. The method according to claim 1, characterized in that, It also includes withdrawal notification and execution steps: After the rights holder initiates withdrawal, the system changes the corresponding scene status of the digital human face in the authorization status matrix to "withdrawn" and broadcasts the withdrawal notification to all connected distribution nodes through the distributed status network; the withdrawal notification includes a withdrawal timestamp, withdrawal scope identifier, effective time, and withdrawal reason code; each application platform receives the withdrawal notification in real time through Webhook push; application platforms without Webhook will query the "withdrawn" status through the distribution node and automatically refuse to call on the next call; after receiving the notification, the application platform must return a confirmation receipt within a specified time limit; platforms that fail to return a receipt will be continuously retried by the system and the non-response event will be recorded; After each platform confirms receipt, the withdrawal process is carried out in a hierarchical manner: stop the generation of new models and destroy original materials, remove unpublished content, and remove published content; verify each uploaded work to ensure it is within the scope of informed consent, keep the works that pass the verification on display, and put those that fail the verification into the removal window period; the withdrawal notice, the confirmation receipts from each platform, and the execution results are all written into the authorization status change log and timestamps are applied for for evidence storage.

10. A multi-layered fingerprint authorization and rights protection management system for real-person digital human faces, characterized in that, The system includes a data collection and filing module, an informed consent module, a filing and distribution module, an authorized invocation module, a fine-tuning and feedback module, an infringement detection module, and a rights protection execution module. The comparison benchmark of the infringement detection module comes from the data collection and filing module's data collection and binding package, and the system verifies the authorization status of the filing and distribution module through a reverse notification. The infringement detection module includes a multi-layer overlay fingerprint analysis unit, used to analyze the fingerprint chain of the model to be detected layer by layer and reconstruct the lineage path. The infringement detection module also includes an import sampling comparison unit, which supports rights holders to upload suspicious images, videos, or model files, automatically extracts facial feature vectors, sub-features, and model fingerprints, compares them side by side with the original files in the data collection and binding package, and outputs a similarity score, a sub-feature difference heatmap, and a fingerprint chain lineage path.

11. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 9.

Images