Partial AI Weight Compression for Secure Memory Execution

Resolve Bottlenecks,
Find Innovative Solutions
Generate 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

VSEngineering 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

Engineering Contradiction:
ImproveAI model securityVSAvoidunauthorized access risk
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

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.

Inventive Principle:
Principle #1Segmentation

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.

Inventive Principle:
Principle #24Intermediary (Mediator)

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

Engineering Contradiction:
Improvedata protectionVSAvoidcompression processing time
Core Design Contradiction:
ReliabilityVSLoss of time

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.

Inventive Principle:
Principle #10Preliminary action

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.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4629090A1Electronic device, method, and non-transitory computer-readable recording medium for protecting artificial intelligence model
Publication Date: 2025.10.08 SAMSUNG ELECTRONICS CO LTD
  • EP4629090A1 patent drawingFigure 1
  • EP4629090A1 patent drawingFigure 2A
  • EP4629090A1 patent drawingFigure 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.