AI Weight Decomposition Across Secure and Unsecured Execution Environments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The proliferation of personalized AI models introduces significant privacy and data security concerns due to the risk of privacy breaches from extensive personal data usage, which existing technologies have not adequately addressed.
Innovation Solution
The method involves decomposing AI model weights into lower-rank matrices, designating sensitive data for processing within a secure execution environment (SEE) and non-sensitive data for processing within an unsecured execution environment (UEE), encrypting and storing sensitive data in secure memory, and performing computations across these environments to enhance security and privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If personalized AI models use extensive personal data to improve model accuracy and personalization, then the model's ability to serve individual users is improved, but the risk of privacy breaches and data security vulnerabilities increases
Solution Approach 1:
The patent segments the AI model weights into three distinct matrices (U, V, and Σ) based on sensitivity classification. The third matrix Σ containing sensitive personal data characteristics is isolated and processed separately in a secure execution environment, while the first and second matrices are processed in the standard execution environment. This segmentation allows the model to maintain personalization capabilities while protecting sensitive data from breaches.
Solution Approach 2:
The patent introduces a secure execution environment as an intermediary between the standard execution environment and the sensitive data processing. This intermediary environment (including secure memory and encrypted communication channels) mediates the handling of sensitive personal data, allowing processing to occur without exposing the data to potential breaches in the standard environment.
2Object-affected harmful factors
If all AI model computations are performed in a secure execution environment to protect sensitive data, then data security is improved, but computational efficiency and processing speed deteriorate
Solution Approach 1:
The patent divides the computational workload by segmenting model weights into sensitive and non-sensitive components. Computations involving the third matrix Σ are restricted to the secure execution environment, while computations involving the first and second matrices can proceed in the standard execution environment at full speed. This segmentation resolves the contradiction by limiting secure processing only where necessary.
Solution Approach 2:
The patent applies local quality by providing enhanced security measures (secure execution environment, encryption, secure memory) specifically for the third matrix Σ that handles sensitive data, while the first and second matrices process data with standard security measures. This localized approach to security maintains high computational efficiency for non-sensitive operations while protecting sensitive data.
3Object-affected harmful factors
If AI model weights are decomposed into multiple matrices and processed across different execution environments, then data security and privacy protection are improved, but system complexity and implementation difficulty increase
Solution Approach 1:
The patent segments the model weights into three matrices (U, V, Σ) where the segmentation is based on the sensitivity of the data characteristics. This mathematical decomposition provides a clear, structured approach to privacy protection that, while increasing some complexity, follows established linear algebra principles and can be integrated systematically into existing AI frameworks.
Solution Approach 2:
The secure execution environment acts as an intermediary layer that manages the complexity of secure processing. It provides standardized interfaces for encrypted data transfer, secure computation, and key management, thereby abstracting away the underlying security complexity from the main system architecture and making implementation more manageable.
4Object-affected harmful factors
If sensitive data is encrypted and stored in secure memory within the execution environment, then data protection against unauthorized access is improved, but data accessibility and retrieval time worsen
Solution Approach 1:
The patent applies preliminary action by encrypting and storing the third matrix Σ in secure memory before it is needed for processing. The secure memory system prepares the encrypted data in advance, and when access is required, the system can quickly retrieve and decrypt the data using stored encryption keys, minimizing retrieval time while maintaining continuous protection against unauthorized access.
Solution Approach 2:
The secure memory system serves as an intermediary between the standard memory system and the sensitive data storage. It provides optimized access mechanisms for encrypted data, including secure key management and efficient encryption/decryption operations, thereby reducing the time penalty associated with secure data retrieval compared to standard unencrypted storage.
Data Source
AI summary
Various embodiments include systems and methods for securing artificial intelligence models in a computing device. Embodiment methods may include decomposing original model weights into lower-rank matrices including a first matrix, a second matrix, and a third matrix. The first matrix and the second matrix may be designated for processing within an unsecured execution environment (UEE). The third matrix (Σ) may be designated for processing within a secure execution environment (SEE). The third matrix (Σ) may be encrypted in the UEE and transferred to the SEE where it may be encrypted. Secure computations to generate inference results may be performed in the SEE, and the inference results or third matrix (Σ) stored in encrypted form in a secure memory within the SEE.