Camouflaged Neural Network Model for Edge AI Security

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In edge computing environments, there is a challenge in ensuring the security and privacy of artificial intelligence models and data transmitted between client devices and edge cloud infrastructures, particularly due to concerns about intellectual property theft and latency issues associated with establishing secure channels.

Innovation Solution

The implementation of a camouflaged neural network model that applies transformations to input and output data at the client device before transmission, using statistical noise and scrambling to hide the model's functionality without encryption, allowing secure transmission over unsecured channels and reducing the need for expensive trust arrangements and computational overhead.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If encryption is used to protect data transmission between client devices and edge cloud infrastructures, then security is improved, but latency increases due to the computational overhead of establishing secure channels

Engineering Contradiction:
ImprovesecurityVSAvoidlatency
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The camouflaging transformation is applied to the neural network model in advance before deployment to the edge infrastructure. This preliminary action embeds the security mechanism directly into the model structure, eliminating the need for runtime encryption/decryption operations and associated latency penalties.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

A camouflaging transformation acts as an intermediary between the original neural network model and the transmission channel. This transformation layer obfuscates the model's functionality and protects against IP theft without requiring traditional encryption protocols, thus avoiding the latency associated with secure channel establishment.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If traditional security measures are implemented to protect AI models from intellectual property theft, then security is improved, but device complexity and computational overhead increase

Engineering Contradiction:
ImprovesecurityVSAvoidcomputational overhead
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The invention changes the parameters of the neural network model by applying a camouflaging transformation that modifies the model's input-output relationship. This transformation alters the model's parameters in a way that protects against reverse engineering while maintaining the model's functionality, avoiding the need for complex security infrastructure.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The camouflaging transformation uses computationally lightweight operations that can be applied once during model preparation. Unlike continuous encryption protocols that require ongoing computational resources, this approach provides security through a single, inexpensive transformation step.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP3547183B1Ai model and data camouflaging techniques for cloud edge
Publication Date: 2024.10.30 INTEL CORP
  • EP3547183B1 patent drawingFigure 1
  • EP3547183B1 patent drawingFigure 2
  • EP3547183B1 patent drawingFigure 3

AI summary

Systems and techniques for AI model and data camouflaging techniques for cloud edge are described herein. In an example, a neural network transformation system is adapted to receive, from a client, camouflaged input data, the camouflaged input data resulting from application of a first encoding transformation to raw input data. The neural network transformation system may be further adapted to use the camouflaged input data as input to a neural network model, the neural network model created using a training data set created by applying the first encoding transformation on training data. The neural network transformation system may be further adapted to receive a result from the neural network model and transmit output data to the client, the output data based on the result.