Transcoding Neural Network for Secured Data Interfaces

Solution Overview

Problem

Conventional neural network architectures lack a practical mechanism for protecting sensitive data, as they are not adaptable to obfuscated data due to a one-to-one mapping between data and meaning, making them vulnerable to security and privacy issues, especially in hosted environments like AWS where neural networks perform computations off-site in the cloud.

Innovation Solution

The implementation of a transcoding neural network that embeds a shared-secret obfuscation mechanism within existing neural networks, using a novel form of transcoding that maps continuous numbers to multidimensional projections, creating a protected neural network that outputs transcoded data, thereby protecting sensitive information at all stages of execution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional neural network architectures are used with one-to-one mapping between data and meaning, then data processing accuracy is maintained, but data security and privacy are compromised

Engineering Contradiction:

Improvedata processing accuracyVSAvoiddata security vulnerability

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent introduces an intermediary encoding layer that transforms clear data into encoded representations before neural network processing. This intermediary layer acts as a mediator between the data and the neural network, allowing accurate processing while preventing direct exposure of sensitive information. The encoded data maintains the necessary structure for computation but obscures the original meaning from unauthorized observers.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the parameter representation of data by transforming it from clear numerical values into encoded forms with specific properties (such as being centered around zero or having specific statistical characteristics). This parameter transformation allows the neural network to process the data accurately while the changed parameters prevent direct interpretation of the original sensitive information.

Inventive Principle:

Principle #35Parameter changes

2Object-affected harmful factors

If data obfuscation methods are applied to protect sensitive information, then data security is improved, but neural network usability deteriorates due to limited mathematical operations

Engineering Contradiction:

Improvedata securityVSAvoidneural network adaptability to obfuscated data

Core Design Contradiction:

Object-affected harmful factorsVSAdaptability or versatility

Solution Approach 1:

The patent changes the parameter representation of data by transforming it from clear numerical values into encoded forms with specific properties (such as being centered around zero or having specific statistical characteristics). This parameter transformation allows the neural network to process the data accurately while the changed parameters prevent direct interpretation of the original sensitive information.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent applies preliminary encoding transformations to the data before it enters the neural network. This preliminary action prepares the data in advance to be compatible with neural network operations while maintaining security. The encoding is designed beforehand to ensure that all necessary mathematical operations can be performed on the encoded representations.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If clear data is used for neural network processing, then model accuracy is maintained, but privacy requirements and security risks increase in hosted environments

Engineering Contradiction:

Improvemodel accuracyVSAvoidprivacy exposure risk

Core Design Contradiction:

ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent introduces an intermediary encoding layer that transforms clear data into encoded representations before neural network processing. This intermediary layer acts as a mediator between the data and the neural network, allowing accurate processing while preventing direct exposure of sensitive information. The encoded data maintains the necessary structure for computation but obscures the original meaning from unauthorized observers.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies preliminary encoding transformations to the data before it enters the neural network. This preliminary action prepares the data in advance to be compatible with neural network operations while maintaining security. The encoding is designed beforehand to ensure that all necessary mathematical operations can be performed on the encoded representations.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20240330513A1System and method for creating secured neural networks

Publication Date: 2024.10.03 IRDETO BV

AI summary

A method of embedding an implementation of a shared-secret obfuscation mechanism into an existing neural network to provide the neural network with more secure data interfaces is disclosed. Disclosed implementations leverage a novel form of transcoding that can be implemented within a conventional neural network. This transcoding maps a span along a continuous number line on to a segment that defines a multidimensional projection. A definition of the transcoding properties (the mapping of spans to segments) is supplied to a form of compiler which produces a transcoding neural network (architecture+weights and biases) that can be embedded within an existing neural to cause the neural network to output transcoded data.