Portrait de-identification method with controllable similarity

Abstract

The application discloses a controllable-similarity portrait de-identification method, and belongs to the technical field of image processing and personal information security. The application creatively proposes a "15%-85% controllable-similarity interval" system by combining an arbitrary stylized image processing technology with a privacy protection preset threshold technology, completely defines full gradient application thresholds of a half portrait to an eight-portrait, and locks recommended daily social threshold and recommended identity authentication threshold. The application establishes a six-step closed-loop process of "original photo input -> core feature processing and grading control -> preset similarity control -> arbitrary stylized rendering -> dynamic anti-restoration verification -> standardized output", and proposes a dynamic anti-restoration verification standard of R<=K*S*G and R<=preset safety threshold through the cooperation of three mechanisms of face feature grading adjustment, preset similarity control and dynamic anti-restoration verification, wherein K is a basic proportion coefficient, and S is a style restoration coefficient. The application supports portrait generation in full style, full form and full angle, realizes three safety effects of "familiar people can be identified, strangers cannot be restored, and AI cannot be reversely trained", and is suitable for all scenes requiring to use a head portrait or perform identity authentication.

Description

Technical Field

[0001] This invention relates to the fields of image processing technology, personal information security protection technology, and digital image generation technology. Specifically, it relates to a method that achieves a precise balance between personal portrait privacy protection and identity recognition through the coordinated use of three mechanisms: hierarchical adjustment of facial features, preset similarity control, and dynamic anti-reconstruction verification. Background Technology

[0002] With the popularization of internet and artificial intelligence technologies, the problem of personal portrait privacy leakage is becoming increasingly serious:

[0003] Risks of using original photos: As profile pictures or identity verification credentials, there is a risk of infringement of portrait rights, malicious collection and misuse of facial information. AI face-swapping and deepfake technologies make publicly available real photos easily exploitable. The drawbacks of purely cartoon icons: While they can protect privacy, they lack personal distinctiveness, cannot establish personal identity tags, and are unsuitable for identity verification scenarios.

[0004] The fundamental flaw of existing stylization techniques is that the generated virtual avatars are either overly realistic (similarity > 80%), leading to privacy leaks, or overly abstract (similarity < 20%), losing personal identity characteristics. A predetermined, controllable balance between "privacy protection" and "identity recognition" cannot be achieved.

[0005] Different styles have inconsistent protection effects: realistic style avatars still have a high recognition rate even with low similarity, while cartoon style avatars do not. Existing technology does not take into account the impact of style on the effectiveness of privacy protection.

[0006] The contradiction in identity authentication scenarios: Access control, payment and other systems usually require 40%-60% similarity to accurately identify, but headshots in this range are very easy to reconstruct using image reconstruction technology.

[0007] Lack of standardized solutions: Existing technologies rely on random output and subjective adjustments to generate effects, making it impossible to apply them in a standardized manner on a large scale.

[0008] In summary, there is currently no effective technological solution worldwide that can simultaneously solve the above problems. Summary of the Invention

[0009] (a) Technical problems to be solved The technical problem to be solved by this invention is: how to achieve preset and controllable portrait similarity of any style and form without modifying the underlying model of existing image generation tools, while meeting the dual requirements of "privacy protection" and "identity recognition", and solving the problem of inconsistent privacy protection effects under different styles.

[0010] (II) Technical Solution To address the aforementioned technical problems, this invention provides a portrait de-identification method with controllable similarity, comprising the following six core actions: Action 1: Raw Photo Input and Preprocessing The system acquires the user's original real-life photo and preprocesses it. This preprocessing includes cropping, brightness adjustment, contrast adjustment, and background removal. The processed image retains a clear facial area, with the face occupying 60%-80% of the total image area.

[0011] The original real-life photos were taken from angles including a frontal eye view, a 45° oblique angle, a pure left / right side profile angle, and a slight tilt angle; the compositions included close-up head shots, head and shoulder half-body shots, bust shots, centered symmetrical shots, and slightly offset shots.

[0012] Action 2: Core Feature Processing and Hierarchical Control Facial feature processing is performed on the preprocessed original photo to achieve the following: Information corresponding to core identification features is retained to the highest degree; Information corresponding to secondary identification features is retained to a lesser extent; Information corresponding to privacy-sensitive features has been completely removed.

[0013] The core recognition features include facial contour, brow bone contour, basic eye socket position, basic nose tip position, and jawline contour; the secondary recognition features include nose details, lip contour, ear contour, and basic hairstyle contour; and the privacy-sensitive features include skin texture, pores, fine wrinkles, spots, scars, hair details, pupil texture, and light and shadow gradient.

[0014] This processing can be achieved through explicit feature grading adjustment rules or implicitly through a trained deep learning model. As long as the processing result satisfies the aforementioned differences in information retention, it falls within the scope of this step.

[0015] One way to implement the above processing is as follows: extract geometric contours as core recognition features through facial key point detection, separate the skin region using a face parsing network, perform frequency domain low-pass filtering on the skin region to remove privacy-sensitive features such as texture details, and then fuse it with the geometric contours.

[0016] It should be noted that the "facial features" mentioned in this invention include not only facial geometric features, but also any other features that can be used for identity recognition, including but not limited to expression features, posture features, clothing features, hairstyle features, and overall visual style features. The "similarity" mentioned in this invention refers to the overall identity similarity based on all the above features.

[0017] Action 3: Preset Similarity Control A preset similarity threshold range is set to control the facial feature matching degree between the generated stylized virtual avatar and the preprocessed original photo to be between 15% and 85%. The lower limit of 15% is determined based on the minimum visual threshold of identity recognizability. When the similarity is below 15%, the generated virtual avatar will lose its association with the original user's identity and will not meet the requirement of "identity recognizability". The upper limit of 85% is determined based on the highest security threshold for privacy protection. When the similarity is above 85%, the virtual avatar retains too many original biometric features, which are not only easily restored by image restoration algorithms, but also visually approach the original photo, thus failing to meet the requirement of "privacy protection".

[0018] The recommended similarity threshold range is as follows: • Recommended extremely high privacy preset threshold: 15% to less than 25% • Recommended daily social interaction threshold: 25% to less than 35% • Recommended default threshold for identity verification: 35% to less than 45% • Recommended preset threshold for semi-public scenarios: 45% to less than 55% • Recommended high recognition threshold: 55% to less than 65% • Recommended high-recognition preset threshold: 65% to less than 75% • Recommended preset threshold for extremely high recognition: 75%-85% Action 4: Arbitrary Stylized Rendering Based on user preferences and usage scenarios, arbitrary stylization parameters can be selected to render the processed facial features, generating a stylized virtual avatar. It supports multiple visual styles, including 2D planar styles, 3D stereoscopic styles, and special effects styles, all based on computer graphics.

[0019] Optionally, privacy enhancement processing is performed during rendering, including at least one of the following: • For facial feature points that are sensitive to the image restoration process, such as the corners of the eyes, the peaks of the eyebrows, the tip of the nose, and the corners of the mouth, perform random position shifts of ±1-3 pixels; • An invisible perturbation noise based on an adversarial example generation algorithm is embedded in the generated virtual image. The perturbation noise is optimized for a deep learning-based image restoration model. When the image restoration model processes the virtual image, the perturbation noise triggers gradient confusion or feature confusion in the restoration model, thereby outputting a blurred or distorted restoration result, thus achieving active defense against the restoration process. • Reduce the detail on one side of non-critical facial features to break the original natural symmetry; • Add low-intensity, invisible random noise.

[0020] It should be noted that the execution order of the third and fourth actions can be changed or executed in parallel as needed.

[0021] Action 5: Dynamic Reverse Verification (Core Step) The generated stylized virtual avatar undergoes dynamic de-restoration verification, which includes the following sub-steps: a) Use at least one image restoration processing algorithm to restore the generated stylized virtual image; b) For all the restoration results output by each algorithm, take the highest similarity value as the restoration similarity value R corresponding to that algorithm; c) Calculate the similarity value G between the generated stylized virtual avatar and the preprocessed original photograph; d) The verification criterion is: the restoration similarity value R corresponding to all the algorithms described satisfies the following condition: · R ≤ K × S × G • R ≤ Preset safety threshold in: • K is the basic scaling factor, ranging from 0.3 to 0.9, reflecting the overall restoration capability of current image restoration technologies. The method for determining the value of K is as follows: it is calibrated based on the highest restoration capability of current mainstream AI restoration algorithms on a standard test set, and the version number is updated with each algorithm iteration. • S is the style restoration coefficient, ranging from 0.2 to 0.8, determined based on the style type of the generated virtual avatar. A larger S value indicates that the style is more easily restored, and the weaker the privacy protection capability. For realistic styles, the value of S ranges from 0.6 to 0.8. For a semi-realistic blending style, the value of S ranges from 0.4 to 0.6. For abstract art styles, the value of S ranges from 0.2 to 0.4. For other visual styles, the value of S ranges from 0.2 to 0.8. • G represents the similarity value between the generated stylized virtual avatar and the original photograph; The preset safety threshold ranges from 20% to 40%.

[0022] The method for determining the style restoration coefficient S is as follows: Under a standard testing environment, current mainstream image restoration algorithms are selected to perform restoration tests on a large sample set of virtual images of this style. The highest restoration rate of the sample set is statistically analyzed and used as the baseline of the inherent attribute of this style against image restoration, which is then independently calibrated to obtain the S value. The S value is independent of the similarity G of the virtual image to be verified, reflecting the inherent resistance of a specific visual style to image restoration algorithms.

[0023] In a preferred embodiment, the base ratio coefficient K is 0.7 and the preset safety threshold is 30%.

[0024] e) If the above verification criteria are met, the verification is successful; if not, return to action three, reduce the current similarity value G within the preset similarity threshold range, and regenerate the stylized virtual image; if the verification still fails after the preset maximum number of retries, output a failure message and suggest that the user adjust the style parameters or select a lower similarity level.

[0025] It should be noted that the above verification formula is an empirical verification rule derived by the applicant through numerous comparative experiments. The applicant found through experiments that, under the current level of mainstream image restoration technology, when K×S≤0.55, the restoration similarity R never exceeds 30%. Therefore, in a preferred embodiment, K×S≤0.55 is recommended as the safety design boundary of the system. Those skilled in the art can make appropriate fine-tuning of the above coefficients according to the actual application scenario and the future development of image restoration technology.

[0026] In practical engineering implementation, currently known mainstream image restoration tools (including but not limited to GFPGAN, CodeFormer, Real-ESRGAN, etc.) can be integrated as verification benchmarks. As image restoration technology develops, the integrated restoration tool library can be updated at any time to ensure the effectiveness of the verification. The scope of protection of this invention is not limited to any specific restoration tool or algorithm.

[0027] Action Six: Standardized Output Outputs verified, stylized virtual avatars. Supports various formats including static digital images, animated avatars, short video avatars, 3D models, real-time driven digital humans, and virtual avatars from virtual and augmented reality.

[0028] When the generated virtual avatar is in the form of animation, video, 3D, or digital human, the dynamic anti-reconstruction verification includes: performing anti-reconstruction verification on each frame of the animation avatar or video; performing anti-reconstruction verification on all perspectives of the 3D model or digital human; until the reconstruction results of all frames and all perspectives meet the verification criteria.

[0029] (III) Beneficial Effects The present invention has the following beneficial effects: 1. Triple security protection: achieving a triple security effect of "familiar people can be identified, strangers cannot be identified, and AI cannot be reverse trained".

[0030] 2. Precise balance: For the first time, it achieves preset controllable similarity within the range of 15%-85%, fully covering application scenarios of the entire gradient from bipartite to octet image.

[0031] 3. Full style coverage: Supports multiple visual styles and quantifies the differences in the difficulty of reproducing different styles through the style reproduction coefficient S, ensuring that the preset privacy and security standards can be met under various styles.

[0032] 4. All formats and angles: Supports all output formats and all shooting angles, including static, dynamic, video, 3D, digital human, VR / AR, etc.

[0033] 5. Multi-layer privacy enhancement: Provides a variety of enhancement techniques such as feature point perturbation, adversarial protection labeling, asymmetric blurring, and random noise.

[0034] 6. Low cost and easy implementation: It can be implemented without modifying the underlying model of existing image generation tools, simply through parameter combination and process design.

[0035] 7. Compliance Support: The method of this invention complies with the core requirements of GB / T 41819-2022 "Information Security Technology - Requirements for Facial Recognition Data Security" regarding data minimization (Article 5.1), non-unnecessary storage of original images (Article 5.2), updatable irreversible transformation (Article 6.2), and de-identification (Article 7.2). It also meets the relevant provisions of GB / T 38671-2020 "Information Security Technology - Technical Requirements for Remote Facial Recognition Systems" and GB / T 35273-2020 "Information Security Technology - Personal Information Security Specification." Through the synergistic effect of facial feature hierarchical adjustment, preset similarity control, and dynamic reverse verification mechanisms, this invention provides quantifiable and verifiable technical means for "de-identification" and "irreversible transformation." Enterprises and platforms adopting this method can use it to demonstrate to regulatory agencies that their avatar display systems meet national standards for personal information protection compliance, effectively reducing the compliance costs of personal information protection. Attached Figure Description

[0036] Figure 1 : The overall flowchart of the method of this invention. It shows the complete process from action one to action six, especially the callback arrow pointing to action three when action five fails verification, and the branch that outputs a failure message after reaching the maximum number of retries.

[0037] Figure 2A diagram illustrating facial feature grading. A simplified facial outline is used to label three levels: core recognition features, secondary recognition features, and privacy-sensitive features.

[0038] Figure 3 A gradient diagram illustrating different preset similarity intervals. Seven recommended intervals, ranging from half-image to eight-image, are marked using a 0%-100% scale.

[0039] Figure 4 : Schematic diagram of dynamic reverse verification principle. It shows the logical flow of "generating virtual image → restoration algorithm processing → calculating R value → comparing with K×S×G → determining pass / fail → callback or output".

[0040] Figure 5 : A diagram illustrating the matching of style reproduction coefficients S under different styles. It shows the correspondence between realistic style S=0.6-0.8, semi-realistic style S=0.4-0.6, and abstract style S=0.2-0.4. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0042] Example 1: Generating three-dimensional profile pictures for everyday social scenarios 1. Original photo input and preprocessing (Action 1): Obtain a high-definition frontal photo of the user, crop it to 512×512 pixels, adjust the brightness and contrast, remove the background, retain the clear facial area, and the face accounts for 70%.

[0043] 2. Core Feature Processing and Hierarchical Control (Action Two): Facial feature processing is performed on the pre-processed photos. After processing, the information retention rate of core recognition features such as facial contour, brow bone contour, and basic eye socket position is above 96%. Information of secondary recognition features such as nose details and lip contour is partially removed, and information of privacy-sensitive features such as skin texture and pores is completely removed.

[0044] 3. Preset similarity control (Action 3): Select the recommended daily social threshold range of 25% to less than 35%, and set the target similarity G to 30%.

[0045] 4. Stylized rendering (Action 4): Select a hand-drawn style for rendering, and perform ±1 pixel random perturbation on sensitive feature points such as the corners of the eyes, the peaks of the eyebrows, the tip of the nose, and the corners of the mouth.

[0046] 5. Dynamic Reverse Reconstruction Verification (Action Five): The style is abstract art, with a style reconstruction coefficient S=0.3. The base proportion coefficient K is set to 0.7. Three image reconstruction algorithms are used for reconstruction, and the reconstruction similarity R values ​​are all below 5%. Verification shows that all algorithms meet the verification criteria of R≤0.7×0.3×0.3=0.063 (i.e., 6.3%) and R≤30%, therefore the verification is successful.

[0047] 6. Standardized Output (Action Six): Generate a 1024×1024 pixel PNG avatar with a transparent background.

[0048] Example 2: Generation of a four-part avatar for identity authentication scenarios 1. Action 1: Obtain a high-resolution frontal photo of the user, with the face accounting for 75% after preprocessing, and the lighting is uniform and unobstructed.

[0049] 2. Action Two: After processing, all core identification features are retained, secondary identification features are retained, and all privacy-sensitive features are removed.

[0050] 3. Action 3: Select the recommended identity authentication threshold range of 35% to less than 45%, and set the target similarity G to 40%.

[0051] 4. Action Four: Select a semi-realistic rendering style and embed a protective marker based on an adversarial example generation algorithm.

[0052] 5. Action Five: This style is a semi-realistic fusion style, with style restoration coefficients S=0.5 and K=0.7. Three image restoration algorithms were used for restoration. The R values ​​of all restoration results were less than 12%, satisfying the verification criteria of R≤0.7×0.5×0.4=0.14 (i.e., 14%) and R≤30%. The verification passed.

[0053] 6. Action Six: Output a standard avatar for use in access control systems or online identity authentication.

[0054] Example 3: Real-time generation of virtual avatars for video calls 1. Acquire user video streams in real time.

[0055] 2. Perform face detection and feature processing on each frame.

[0056] 3. Control the similarity between the virtual avatar and the original human face to be between 15% and 85%.

[0057] 4. The virtual avatar is driven in real time based on the user's facial expressions and movements.

[0058] 5. Perform dynamic de-reconstruction verification on the generated virtual avatar to ensure that all frames pass verification; if three consecutive frames fail verification, the similarity G is automatically reduced and the avatar is regenerated; if the maximum number of retries is reached and the avatar still fails to pass, the user is prompted to switch styles or reduce the similarity level.

[0059] 6. Output real-time video stream.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A portrait de-identification method with controllable similarity, characterized in that, Includes the following actions: Action 1: Obtain the user's original real-life photo and preprocess it; Action 2: Perform facial feature processing on the preprocessed original photo so that the information corresponding to the core recognition features is retained to a higher degree than the information corresponding to the secondary recognition features, while the information corresponding to privacy-sensitive features is removed. Action 3: Set a preset similarity threshold range to control the matching degree of facial features between the generated stylized virtual image and the preprocessed original photo to be between 15% and 85%; Action 4: Based on the user's personal preferences and usage scenario, select any stylization parameters to render the facial features processed in Action 2, generating a stylized virtual image; Action 5: Perform dynamic de-restoration verification on the generated stylized virtual avatar, the dynamic de-restoration verification including: a) Use at least one image restoration processing algorithm to restore the generated stylized virtual image; b) For all the restoration results output by each algorithm, take the highest similarity value as the restoration similarity value R corresponding to that algorithm; c) Calculate the similarity value G between the generated stylized virtual avatar and the preprocessed original photograph; d) The verification criterion is as follows: the restoration similarity value R corresponding to all the algorithms satisfies R≤K×S×G, and R≤preset safety threshold; where K is the basic proportional coefficient, with a value range of 0.3-0.9; S is the style restoration coefficient, which is pre-calibrated according to the style type of the generated virtual image, with a value range of 0.2-0.

8. The larger the S value, the easier it is for the style to be restored; the preset safety threshold has a value range of 20%-40%. e) If the above verification criteria are met, the verification passes; if not, return to action three, reduce the current similarity value G within the preset similarity threshold range, and regenerate the stylized virtual image; if the verification still fails after the preset maximum number of retries, output a failure message. Action Six: Output the stylized virtual image that has passed the dynamic de-reconstruction verification; The execution order of action three and action four can be changed or executed in parallel as needed.

2. The method according to claim 1, characterized in that, In Action 1, the preprocessing includes cropping, brightness adjustment, contrast adjustment, and background removal. After processing, a clear facial area is retained, with the face accounting for 60%-80% of the total image area.

3. The method according to claim 1, characterized in that, In Action 2, the core recognition features include facial contour, brow bone contour, basic eye socket position, basic nose tip position, and jawline contour; the secondary recognition features include nose details, lip contour, ear contour, and basic hairstyle contour; and the privacy-sensitive features include skin texture, pores, fine wrinkles, spots, scars, hair details, pupil texture, and light and shadow gradient.

4. The method according to claim 1, characterized in that, In Action 2, one way to implement the facial feature processing is as follows: extract geometric contours as core recognition features through facial key point detection, separate the skin region using a face parsing network, perform frequency domain low-pass filtering on the skin region to remove privacy-sensitive features such as texture details, and then fuse it with the geometric contours.

5. The method according to claim 1, characterized in that, In action three, the preset similarity threshold range includes the following recommended threshold ranges: (1) The recommended extremely high privacy preset threshold is 15% to less than 25%; (2) The recommended daily social interaction threshold is 25% to less than 35%; (3) The recommended preset threshold for identity authentication is 35% to less than 45%; (4) The recommended preset threshold for semi-public scenarios is 45% to less than 55%; (5) The recommended higher recognition threshold is 55% to less than 65%; (6) The recommended high recognition threshold is 65% to less than 75%; (7) The recommended threshold for extremely high recognition is 75%-85%.

6. The method according to claim 1, characterized in that, In a preferred embodiment, the base scaling factor K is 0.7; when K is 0.7, it is recommended that the style restoration factor S be in the range of S≤0.

7.

7. The method according to claim 1, characterized in that, The preferred value of the basic proportional coefficient K is 0.7, and the preferred value of the preset safety threshold is 30%.

8. The method according to claim 1 or 7, characterized in that, The value of the style restoration coefficient S is determined based on the style type of the generated virtual avatar: (1) For realistic style, the value of S ranges from 0.6 to 0.8; (2) For the semi-realistic blending style, the value of S ranges from 0.4 to 0.6; (3) For abstract art style, the value of S ranges from 0.2 to 0.4; (4) For other visual styles, the value of S ranges from 0.2 to 0.

8.

9. The method according to claim 1, characterized in that, In Action Four, the arbitrary stylization parameters cover visual expression styles generated based on computer graphics, such as two-dimensional planar style, three-dimensional stereo style, or special effects style.

10. The method according to claim 1, characterized in that, Action four also includes privacy enhancement processing, which includes at least one of the following: a) Randomly offset the facial feature points at the corners of the eyes, brow peaks, nose tips, and corners of the mouth by ±1-3 pixels; b) Embed invisible perturbation noise based on an adversarial sample generation algorithm into the generated virtual image. The perturbation noise is optimized for a deep learning-based image restoration model. When the image restoration model processes the virtual image, the perturbation noise triggers gradient confusion or feature confusion in the restoration model, thereby outputting a blurred or distorted restoration result. c) Reduce the details on one side of non-critical facial features to break the original natural symmetry; d) Add low-intensity, invisible random noise.

11. The method according to claim 1, characterized in that, The generated virtual avatars include static digital images, dynamic avatars, short video avatars, 3D models, real-time driven digital humans, and virtual avatars in virtual reality and augmented reality. When the generated virtual image is in the form of animation, video, 3D, or digital human, the dynamic anti-reconstruction verification includes: performing anti-reconstruction verification on each frame of the dynamic avatar or video; Perform reverse reconstruction verification on all perspectives of the 3D model or digital human.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-11.

13. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-11.

Images