Partial AI Weight Compression for Secure Memory Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in protecting artificial intelligence models from unauthorized access and reconstruction by malicious users, particularly when data related to these models is stored in volatile and non-volatile memory.
Innovation Solution
Implementing a compressor and decompressor within an electronic device to compress and decompress partial weights of an AI model based on specific algorithms, allowing secure storage and execution of these weights in volatile memory while minimizing data exposure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI model weights are stored in volatile and non-volatile memory, then the AI model can be executed and stored, but the system becomes vulnerable to unauthorized access and data theft
Solution Approach 1:
The patent divides AI model weights into multiple segments and stores them in different memory locations. The compressor breaks down the original weights into compressed partial weights, which are then distributed across volatile and non-volatile memory. This segmentation prevents malicious users from accessing complete model information, thereby enhancing security while maintaining functionality.
Solution Approach 2:
The patent introduces a compressor and decompressor as intermediary components between the AI model weights and the memory system. These components encrypt and compress the weights before storage, acting as a protective barrier. The decompressor then reconstructs the weights only when needed for execution, preventing direct access to uncompressed model data.
2Reliability
If compressed partial weights are loaded into volatile memory for execution, then the AI operation can be performed securely, but additional compression and decompression processing time is required
Solution Approach 1:
The patent applies compression to AI model weights in advance during the model loading phase, before execution begins. The compressor pre-processes the weights and stores them in compressed form in non-volatile memory. This preliminary compression eliminates the need for real-time compression during execution, reducing latency and improving operational efficiency.
Solution Approach 2:
The patent replaces traditional uncompressed weight storage and transmission mechanisms with a compression-based system. Instead of directly loading full weights into volatile memory, the system uses compression algorithms to reduce data size and decompression circuits to efficiently reconstruct weights only when needed, substituting mechanical data movement with more efficient compressed data handling.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
An electronic device is disclosed. The electronic device stores weights related to an artificial intelligence (AI) model in a non-volatile memory. The electronic device, based on a load request for a specified operation, compresses, through a compressor, partial weights related to the specified operation among the weights stored in the non-volatile memory, and loads the compressed partial weights into a first volatile memory. The electronic device decompresses, through a decompressor, the partial weights loaded into the first volatile memory with the compressed state and loads the decompressed partial weights into a second volatile memory. The electronic device performs the specified operation based on the partial weights with the decompressed state, loaded into the second volatile memory.