Confidential Computation-Based Electronic Medical Record Quality Control Model Algorithm

Abstract

This invention discloses an electronic medical record (EMR) quality control model algorithm based on confidential computing, belonging to the fields of information security and artificial intelligence technology. By introducing Intel SGX (Software Guard Extensions) technology and combining it with the OpenVINO security plugin of Gramine-SGX, the algorithm achieves secure deployment and efficient inference of the EMR quality control model. This invention provides a highly adaptable and accurate EMR model quality control algorithm to meet the increasing demands of healthcare, while also providing security guarantees for model prediction AI model inference acceleration, model storage, and model inference computation. It offers an efficient and compatible solution for the protection of AI model intellectual property rights and the acceleration of model inference involved in current information-based healthcare algorithms.

Description

Technical Field

[0001] This invention discloses an electronic medical record quality control model algorithm based on confidential computing, belonging to the fields of information security and artificial intelligence technology. Background Technology

[0002] With the rapid development of medical technology and artificial intelligence, traditional electronic medical record (EMR) quality control algorithms are struggling to adapt to the complex and ever-changing medical needs. Diverse and complex EMR quality assessment algorithms are increasingly being combined with AI models for efficient implementation. However, the growing number of AI models stored and running on third-party cloud computing servers raises risks of data breaches, intellectual property infringement, and security attacks. Summary of the Invention

[0003] 1. Technical problems to be solved

[0004] (1) Existing related technologies:

[0005] CN118445280A - Artificial Intelligence-Based Quality Control Method for Electronic Medical Records

[0006] CN118350017A - A Secure and Trusted Use Method and System for Large Models Based on Confidential Computation

[0007] CN117668887A - A customizable balance between security and performance in homomorphic encryption based on confidential computation.

[0008] (2) The following problems and shortcomings also exist in the background technology:

[0009] 1) Traditional quality control algorithms have poor rule adaptability and low recognition accuracy.

[0010] Typically based on fixed rules and predefined logic, these rules are often simple and struggle to adapt to the ever-increasing new medical needs and changes, while also failing to identify complex patterns and anomalies.

[0011] 2) Difficulty in capturing relationships between complex medical data

[0012] Traditional rules struggle to fully capture the complexity of medical data and the relationships between multiple variables.

[0013] 3) Privacy leaks in quality control models

[0014] The models in electronic medical record quality control algorithms typically use sensitive datasets from medical institutions. If the models are not encrypted, they are very vulnerable to attacks, leading to model leakage and the risk of the results being stolen.

[0015] 4) Transparent reasoning and calculation of the quality control model

[0016] Attackers can physically access the model's output (e.g., classification or prediction results) and input data in memory, and use reverse engineering techniques to extract features from the training data or even the raw data.

[0017] 5) The quality control model cannot be centrally managed and optimized.

[0018] Complex electronic medical record quality control algorithms often require multiple models to work together. The more models there are, the higher the demands on the current server, making it impossible to distribute computing power effectively, resulting in a sharp drop in access speed. Maintenance and optimization costs also increase.

[0019] (3) The cause of the problem:

[0020] 1) Traditional quality control rules typically rely on predefined rules and standards, using logical judgments to detect the accuracy and consistency of data. This means that updating the rules is time-consuming and labor-intensive when medical practices, regulations, or technological advancements change. Furthermore, the simplicity of the rules can easily lead to misjudgments or omissions, especially in complex situations, failing to fully identify subtle differences in the data.

[0021] 2) Traditional quality control algorithms are rule-driven, while AI model-based quality control algorithms are data-driven. The complexity of relationships between data is less than the complexity of relationships modeled through data in AI.

[0022] 3) The AI ​​industry is booming, but there is a lack of standards and best practices for encryption models in the medical field. At the same time, encryption is expensive, complex to implement, and has low performance in practical applications, which prevents it from being widely used.

[0023] 4) Among existing privacy computing technologies, such as homomorphic encryption, the use of ciphertext in computation and the output of ciphertext results result in slow AI model predictions, making it impossible to port this technology to the AI ​​field.

[0024] The models used in electronic medical record quality control algorithms are often very different. Different training algorithms, training frameworks and data processing methods, without a unified and standardized interface, make centralized management and optimization difficult and complex.

[0025] To address the aforementioned issues, this invention discloses an electronic medical record quality control model algorithm based on confidential computing. It employs the OpenVINO security plugin of Gramine-SGX and introduces IntelSGX (Software Guard Extensions) technology to implement the electronic medical record quality control model algorithm based on confidential computing.

[0026] 2. Technical Solution

[0027] This invention provides a highly adaptable and accurate electronic medical record model quality control algorithm to meet the ever-increasing medical needs, while also providing security guarantees for model prediction AI model inference acceleration, model storage, and model inference computation. It offers an efficient and compatible solution for the protection of AI model intellectual property rights and the acceleration of model inference involved in current information-based medical algorithms.

[0028] (1) Rule adaptability and recognition accuracy have been greatly improved.

[0029] Unlike traditional quality control algorithms, which suffer from high costs and low accuracy due to rule changes, AI quality control models can be retrained with new data and used to replace existing models, thus adapting to changing medical needs, improving adaptability, and reducing costs. Furthermore, training with large amounts of data enables the identification of complex patterns and anomalies, reducing the probability of misjudgments and missed diagnoses, and improving accuracy.

[0030] (2) Effectively capture complex relationships between data

[0031] Unlike traditional quality control algorithms, which cannot understand complex data relationships, AI quality control models can effectively identify and model complex relationships through machine learning algorithms, helping to discover potential medical problems.

[0032] (3) Improved AI model inference performance

[0033] While existing homomorphic confidentiality techniques protect data during storage and computation, they suffer from high computational overhead and low performance. Introducing SGX effectively addresses this deficiency because data is processed in plaintext within confidential memory; encryption and decryption operations primarily occur during data transmission and storage, resulting in computation speeds approaching those of unencrypted processing.

[0034] (4) Intellectual Property Protection of AI Models

[0035] Both SGX and homomorphic encryption are technologies used to protect data privacy and intellectual property. However, homomorphic encryption cannot effectively balance AI model performance and intellectual property protection, making it far less suitable than using SGX for AI intellectual property protection and product deployment.

[0036] (5) Centralized Model Management and Optimization

[0037] In existing management methods, as the complexity of electronic medical record quality control algorithms increases, the optimization and maintenance costs also rise accordingly, failing to meet current needs. Using the OpenVINO inference acceleration framework, the provided AI model server can effectively manage and optimize multiple AI models, and can access models for prediction via gRPC or HTTP REST, effectively distributing computing power and laying the foundation for AI model expansion.

[0038] 3. Beneficial effects

[0039] (1) Significantly improved the inference speed of AI models

[0040] (2) Provides reliable intellectual property protection for AI models.

[0041] (3) While protecting the intellectual property rights of AI models, it also takes into account the improvement of the reasoning performance of AI models. Attached Figure Description

[0042] none. Detailed Implementation

[0043] The following are the implementation steps of the electronic medical record quality control model algorithm based on confidential computing:

[0044] (1) Data preparation

[0045] Prepare the training models required by the algorithm provided by the medical institution, including the AI ​​model itself and the model input feature data.

[0046] (2) Standardized model format

[0047] Different communities have different model formats (spaCy, Hugging Face, etc.). To optimize and deploy them, they must be unified into a consistent format. All model formats will be converted into an intermediate representation based on the OpenVINO model acceleration framework, commonly known as IR (Intermediate Representation).

[0048] (3) Model Optimization

[0049] After the format is converted into a unified representation format, the model needs to be optimized:

[0050] 1) Quantification

[0051] Convert the floating-point model to a fixed-point model (such as INT8) to reduce model size and improve inference speed.

[0052] 2) Pruning

[0053] Remove unimportant weights and neurons to reduce model complexity and improve inference speed.

[0054] 3) Layer fusion

[0055] Multiple layers can be merged into one to reduce memory bandwidth and computation time.

[0056] (4) Environment setup

[0057] Install the specified operating system Ubuntu 20.04 and upgrade the kernel to version 5.11. Install the SGX driver, pull the relevant component code of the Gramine-SGX OpenVINO security plugin from GitHub, set the relevant configuration, execute the compilation instructions, and execute the installation script.

[0058] (5) Model Deployment

[0059] 1) Start the certificate server

[0060] 2) Use the command to generate the relevant model key (isv_keystore.csr.crt) for each model.

[0061] 3) Encrypt the model using the command model key and generate an encrypted model (model.dat) and a model access license (model.masterlic).

[0062] 4) Use the command to specify the certificate server address and model key to generate a temporary model access license (30daylicense.config), and ensure its legality and integrity.

[0063] 5) Use commands to generate customer keys for different types of customers (isv_keystore.csr.crt)

[0064] 6) Use the command to specify the model access license, client key, temporary model access license, and model keystore to generate the client access license (model.lic).

[0065] 7) Use a script to update the customer access license and customer key in the license server's database.

[0066] 8) Configure the model server's configuration file and start the server.

[0067] (6) Remote access

[0068] Use the RESTful API provided by OpenVINO to replace the prediction code in the algorithm's model. Then you can access the algorithm by calling it.

Claims

1. An electronic medical record quality control model algorithm based on confidential computing, characterized in that, The steps include: Step 1: Introduce Intel SGX (Software Guard Extensions) technology, combined with the OpenVINO security plugin of Gramine-SGX; Step 2: Deploy and run the electronic medical record quality control model in the SGX security environment; Step 3: Achieve efficient management and optimization of the model through the OpenVINO inference acceleration framework.

2. The electronic medical record quality control model algorithm based on confidential computing according to claim 1, characterized in that, Step 1 also includes: using AI model training methods to improve the adaptability and recognition accuracy of quality control rules; protecting model intellectual property rights through SGX technology; and achieving security and privacy protection for model inference computation.

3. The electronic medical record quality control model algorithm based on confidential computing according to claim 1, characterized in that, Step 2 addresses the following technical issues: improving rule adaptability and recognition accuracy; effectively capturing relationships between complex medical data; preventing model privacy leaks; protecting the model inference and computation process; and enabling centralized management and optimization of the model.

4. The electronic medical record quality control model algorithm based on confidential computing according to claim 1, characterized in that, The OpenVINO inference acceleration framework in step 3 includes: an AI model server functional module; a gRPC or HTTP REST interface module for accessing model prediction; and a computing power distribution and model expansion module.

5. The electronic medical record quality control model algorithm based on confidential computing according to claim 1, characterized in that, The algorithm is applied in the field of medical informatics in the following ways: a quality control module for electronic medical records; an intellectual property protection module for AI models; and a secure and reliable model deployment module.