CN-121980578-A - Attitude data privacy processing system and method for depth information encryption and identity desensitization

CN121980578ACN 121980578 ACN121980578 ACN 121980578ACN-121980578-A

Abstract

The application belongs to the technical field of data security and privacy protection, and particularly relates to a gesture data privacy processing system and method based on deep information encryption and identity desensitization. The method is executed by end-side equipment and comprises the steps of collecting depth image data of a human body, carrying out normalization processing on the depth image data, carrying out feature extraction and identity desensitization processing on the normalized depth image data by using a lightweight model, wherein the identity desensitization processing comprises weakening identity information in features through an anti-learning mechanism and mapping the features to a desensitization subspace through feature projection, the lightweight model is trained by the cloud platform, carrying out encryption operation on the feature data after desensitization to generate encrypted data, and transmitting the encrypted data to the cloud platform. The application provides a gesture data privacy processing method which has the advantages of full life cycle privacy protection, complete stripping of identity information and high-efficiency operation at the end side.

Inventors

ZHU XIANG
XUE HENG
HUANG PANPAN

Assignees

北京连屏科技股份有限公司

Dates

Publication Date: 20260505
Application Date: 20251213

Claims (10)

1. The utility model provides a gesture data privacy processing system based on degree of depth information encryption and identity desensitization, includes terminal side equipment and high in the clouds platform, its characterized in that, terminal side equipment with the high in the clouds platform passes through safe communication channel to be connected, and: the end side device is used for: Collecting depth image data of a human body; Normalizing the depth image data; Performing feature extraction and identity desensitization on the normalized depth image data by using a lightweight model, wherein the identity desensitization comprises weakening identity information in features by an anti-learning mechanism and mapping the features to a desensitization subspace by feature projection, and the lightweight model is trained by the cloud platform; performing encryption operation on the desensitized characteristic data, generating encrypted data and transmitting the encrypted data to a cloud platform through the secure communication channel; the cloud platform is used for: receiving the encrypted data through the secure communication channel and decrypting to recover the desensitization feature; and performing attitude estimation based on the decrypted desensitized features, and outputting attitude key point coordinates.
2. The system of claim 1, wherein the process of establishing the secure communication channel comprises: When the terminal side equipment is started for the first time, after the integrity of a trusted execution environment is verified through a secure starting chain, an authentication request is sent to the cloud platform through a TLS 1.3 encryption channel, and the authentication request comprises a factory serial number of the terminal side equipment; After the validity of the serial number is checked by the cloud platform, an equipment exclusive authorization certificate is generated, wherein the certificate comprises the serial number, a hash value of a cloud root key, a key derivation algorithm identification KDF2 and a cloud digital signature, and the serial number, the hash value of the cloud root key, the key derivation algorithm identification KDF2 and the cloud digital signature are issued to the terminal side equipment through the TLS 1.3 channel; and after verifying the signature validity of the certificate and confirming that the serial numbers are consistent, the terminal side equipment derives an encryption key through a KDF2 algorithm based on the hash value, the serial numbers and the hardware-cured fixed random number, and stores the encryption key in a trusted area of a hardware security module.
3. The system according to claim 1, wherein: The cloud platform derives a decryption key based on the terminal side equipment serial number and the cloud root key, and performs decryption and integrity check on the received encrypted data.
4. The system according to claim 1, wherein: The volume of the lightweight model is optimized to be smaller than a first preset volume threshold value to adapt to the end-side equipment resource constraint, and The parameter quantity of the desensitization sub-network is smaller than a first preset parameter quantity threshold value so as to ensure the identity desensitization efficiency.
5. A method for processing gesture data privacy based on depth information encryption and identity desensitization, which is executed by an end-side device, and is characterized by comprising the following steps: Collecting depth image data of a human body; Normalizing the depth image data; Performing feature extraction and identity desensitization on the normalized depth image data by using a lightweight model, wherein the identity desensitization comprises weakening identity information in features by an anti-learning mechanism and mapping the features to a desensitization subspace by feature projection, and the lightweight model is trained by the cloud platform; Performing encryption operation on the desensitized characteristic data to generate encrypted data; and transmitting the encrypted data to a cloud platform.
6. The method according to claim 5, wherein: the identity desensitization processing and encryption operation are completed in a trusted execution environment of the end-side device.
7. The method of claim 5, wherein the weakening the identity information in the feature by an anti-learning mechanism comprises: calling a lightweight identity discrimination network deployed at the end side to identify an identity tag from the feature map; And the terminal-side desensitization sub-network is generated according to the judging result adjusting characteristics of the identity judging network until the identity recognition accuracy is lower than a first preset accuracy threshold.
8. The method of claim 5, wherein the mapping features to desensitized subspaces by feature projection comprises: And linearly projecting the features through the projection matrix, and enabling the mutual information of the projected features and the identity tag to be lower than a first preset mutual information threshold value.
9. The attitude data privacy processing method based on depth information encryption and identity desensitization is executed by a cloud platform and is characterized by comprising the following steps of: Receiving encrypted data sent by end-side equipment, wherein the encrypted data is generated by the end-side equipment after carrying out identity desensitization processing on the depth image data, and the identity desensitization processing comprises weakening identity information through antagonism learning and mapping features to a desensitization subspace through feature projection; decrypting the encrypted data to restore the desensitization characteristic; and performing attitude estimation based on the decrypted desensitized features, and outputting attitude key point coordinates.
10. The method of claim 9, wherein outputting the pose keypoint coordinates is preceded by: Adding noise to the coordinates to prevent the reverse inference of the identity of the individual through the output result based on a differential privacy principle, wherein the noise amplitude is dynamically adjusted according to a preset privacy budget and encryption disturbance; And digital signature and audit log recording are performed on the noise added data to ensure the integrity of the output data and the traceability of the operation.

Description

Attitude data privacy processing system and method for depth information encryption and identity desensitization Technical Field The application belongs to the technical field of data security and privacy protection, and particularly relates to a gesture data privacy processing system and method based on deep information encryption and identity desensitization. Background The human body posture recognition technology aims at detecting, tracking and reconstructing joint key points and skeleton structures of a human body in real time by analyzing images or video sequences so as to understand the posture, the action and the behavior of the human body. The technology has become a core driver for numerous applications such as intelligent monitoring, man-machine interaction, rehabilitation medical evaluation and the like. However, as the application scene of the device is continuously expanded to privacy sensitive fields such as medical treatment, home furnishing and the like, a core contradiction is increasingly highlighted, namely, how to ensure high-precision gesture recognition and fully and effectively privacy protection of sensitive data such as related personal biological characteristics and the like. While the prior art attempts to solve this contradiction from different paths, there are significant drawbacks: the risk of privacy disclosure is high: Conventional gesture recognition systems use RGB video or images as input, including rich visual features such as human faces, body contours, clothing textures, etc., which may be further analyzed for identification or personal behavioral feature recognition, thereby creating a greater risk of privacy disclosure. Desensitization is not protected in place: Traditional desensitization modes (including blurred faces, shielding, skeleton substitution and the like) only can weaken explicit features of an image layer, cannot eliminate identity relevance of the feature layer, and cannot meet strict privacy standards. Privacy preserving algorithms are difficult to run on end-side devices: the existing strong privacy protection algorithm (such as like encryption, differential privacy, federal learning, multiparty security calculation and the like) can theoretically improve the data security, but generally has the problems of high calculation complexity, large model reasoning delay and the like, so that the feasibility of the system in actual deployment is low. Disclosure of Invention The embodiment of the application provides a gesture data privacy processing scheme based on depth information encryption and identity desensitization, and aims to solve the three core problems of high privacy disclosure risk, unsensitivity protection in place and poor end side deployment feasibility in the prior art. The first aspect of the embodiment of the application provides a gesture data privacy processing system based on depth information encryption and identity desensitization, which comprises end-side equipment and a cloud platform, wherein the end-side equipment is connected with the cloud platform through a secure communication channel, and: the end side device is used for: Collecting depth image data of a human body; Normalizing the depth image data; Performing feature extraction and identity desensitization on the normalized depth image data by using a lightweight model, wherein the identity desensitization comprises weakening identity information in features by an anti-learning mechanism and mapping the features to a desensitization subspace by feature projection, and the lightweight model is trained by the cloud platform; performing encryption operation on the desensitized characteristic data, generating encrypted data and transmitting the encrypted data to a cloud platform through the secure communication channel; the cloud platform is used for: receiving the encrypted data through the secure communication channel and decrypting to recover the desensitization feature; and performing attitude estimation based on the decrypted desensitized features, and outputting attitude key point coordinates. In some embodiments of the present application, the process of establishing the secure communication channel includes: When the terminal side equipment is started for the first time, after the integrity of a trusted execution environment is verified through a secure starting chain, an authentication request is sent to the cloud platform through a TLS 1.3 encryption channel, and the authentication request comprises a factory serial number of the terminal side equipment; After the validity of the serial number is checked by the cloud platform, an equipment exclusive authorization certificate is generated, wherein the certificate comprises the serial number, a hash value of a cloud root key, a key derivation algorithm identification KDF2 and a cloud digital signature, and the serial number, the hash value of the cloud root key, the key derivation algorithm identification KDF2 and the cl