Machine Learning Processing with Risk-Based Memory Isolation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing concealment information processing technologies, such as those described in US 2020/0082259 A, fail to prevent the leakage of confidential information during deep learning processing, particularly when input information includes personal privacy or trade secrets, as intermediate data is not adequately protected by the enclave.
Innovation Solution
An information processing apparatus with a processor and memory that includes a normal region and an isolation region, where the isolation region holds input data and parameters, and the processor performs semantic determination processing to assess risk before executing layers of machine learning, ensuring sensitive data is processed only in the isolation region when necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional concealment information processing technology is used to keep data encrypted during processing, then confidentiality of personal information is protected, but the degree of freedom of processing is limited to basic computation only
Solution Approach 1:
The patent divides the machine learning model into multiple processing layers and segments the processing workflow into distinct phases: encryption layer, processing layer, and decryption layer. This allows different types of processing operations to be performed at different stages, enabling both basic computation (encrypted) and advanced statistical processing (decrypted) to coexist within the same system.
Solution Approach 2:
The patent performs preliminary decryption of the machine learning model parameters before processing, allowing advanced statistical processing and machine learning operations to be executed on the decrypted data. After processing, the results are re-encrypted before output. This preliminary action enables complex computations that would otherwise be impossible with encrypted data alone.
2Adaptability or versatility
If homomorphic encryption technology is used to enable processing without encryption limits, then degree of freedom of processing is improved, but practical processing speed becomes unachievable
Solution Approach 1:
The patent segments the processing operations into two categories: basic computation operations that can be performed on encrypted data (maintaining security and speed), and advanced statistical processing operations that require decrypted data (enabling versatility). This segmentation allows the system to use fast encrypted processing for routine operations while reserving slower decrypted processing for complex analytical tasks.
Solution Approach 2:
The patent applies decryption partially - only to the extent necessary for advanced processing operations. The machine learning model parameters are decrypted only when advanced statistical processing is required, while basic computations continue to use encrypted data. This partial decryption approach balances the need for processing versatility with the constraint of processing speed.
3Reliability
If TEE function is used to provide a trust region for decrypting and processing encrypted data, then advanced processing is enabled with safety, but intermediate data may still be exposed to leakage risk
Solution Approach 1:
The patent performs preliminary semantic determination processing on the input data before it enters the TEE trust region. This preliminary action identifies and flags potentially sensitive intermediate data, allowing the system to apply additional protective measures or adjust processing parameters to minimize leakage risk before the data is decrypted and processed within the TEE environment.
Solution Approach 2:
The patent implements feedback mechanisms that monitor the processing of intermediate data within the TEE trust region. The semantic determination processing continuously assesses the sensitivity of intermediate results, and based on this feedback, the system can dynamically adjust the level of protection applied to different data elements, ensuring that highly sensitive intermediate data receives enhanced protection while less sensitive data can be processed more efficiently.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
An information processing apparatus that executes machine learning by a plurality of processing layers includes a processor and a memory, a normal region of the memory holds a parameter of the machine learning, an isolation region isolated from the normal region of the memory holds input data to an i-th layer included in the plurality of processing layers and a parameter of the machine learning, the processor determines whether there is a risk in executing processing of the i-th layer for the input data in the normal region based on a content of secret information of the input data in the isolation region, when determining that there is the risk, executes the processing of the i-th layer for the input data in the isolation region, and when determining that there is no risk, executes the processing of the i-th layer for the input data in the normal region.