AI Model Layer Randomization for Secure Transfer

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current cryptographic key-based solutions for securing AI model transfers between host systems and data processing accelerators are slow and impractical, and existing communication channels are vulnerable to data sniffing attacks, especially during AI model training and inference processes.

Innovation Solution

Implementing a randomization module in the host device to randomize AI model layers and memory addresses before transfer to a data processing accelerator, creating a secure processing system that obscures communication channels through layer reordering and memory address randomization, thereby increasing the difficulty for attackers to rebuild the model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If cryptographic key-based solutions are used to encrypt data over communication channels, then security is improved, but processing speed deteriorates and hardware complexity increases

Engineering Contradiction:

ImprovesecurityVSAvoidprocessing speed

Core Design Contradiction:

ReliabilityVSSpeed

Solution Approach 1:

The patent replaces cryptographic key-based encryption (a complex mechanical/systematic process requiring hardware engines) with a statistical randomization approach. By shuffling layer indices and randomizing memory addresses, the system achieves security through obscurity and complexity rather than cryptographic mathematics, thereby eliminating the need for hardware-based cryptographic engines and improving processing speed.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the security approach from cryptographic parameters (keys, algorithms) to statistical parameters (randomization seeds, shuffle operations). The host generates random seeds to shuffle layer indices and randomize memory addresses, transforming the security mechanism from deterministic encryption to probabilistic obfuscation, which is faster and requires less hardware complexity.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If cryptographic key-based solutions are used to encrypt data over communication channels, then security is improved, but device complexity increases

Engineering Contradiction:

ImprovesecurityVSAvoidhardware complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Speed

If data transmission is performed over communication channels without protection, then processing speed is maintained, but security deteriorates making data vulnerable to sniffing attacks

Engineering Contradiction:

Improveprocessing speedVSAvoidsecurity

Core Design Contradiction:

SpeedVSReliability

Solution Approach 1:

The patent replaces cryptographic key-based encryption (a complex mechanical/systematic process requiring hardware engines) with a statistical randomization approach. By shuffling layer indices and randomizing memory addresses, the system achieves security through obfuscation rather than cryptographic mathematics, thereby eliminating the need for hardware-based cryptographic engines while maintaining fast processing speeds.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces an intermediary randomization layer between the host and DP accelerator. The host shuffles layer indices and randomizes memory addresses before transmission, and the DP accelerator reverses these transformations using stored seeds. This intermediary randomization process protects data during transmission without requiring slow cryptographic encryption, maintaining processing speed while improving security.

Inventive Principle:

Principle #24Intermediary (Mediator)

4Productivity

If AI model layers are transferred in organized order, then processing efficiency is improved, but security deteriorates making it easier for attackers to reconstruct the model

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidsecurity

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary randomization to the AI model layers before transfer. The host shuffles layer indices and randomizes memory addresses before transmission to the DP accelerator. This preliminary action obscures the organized structure of the model during transmission, preventing attackers from easily reconstructing the model even if they intercept the data, while the DP accelerator reverses the randomization using stored seeds to maintain processing efficiency.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11657332B2Method for AI model transferring with layer randomization

Publication Date: 2023.05.23 BAIDU USA LLC

AI summary

A method to transfer an artificial intelligence (AI) model includes identifying a plurality of layers of the AI model, the plurality of layers organized in a first ordered list. The method further includes randomizing the plurality of layers by reorganizing the first ordered list into a second ordered list, and transferring the plurality of layers of the AI model to a data processing accelerator in an order defined by the second ordered list.