Neural Network Security via Auto-Encoder Noise Injection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks are vulnerable to adversarial attacks, where imperceptible changes in input data significantly alter the output, and existing security measures either complicate the network unnecessarily or are limited in applicability.

Innovation Solution

Integrate an auto-encoder neural network into the original network to add parasitic noise to the input, specifically trained to introduce high-frequency noise that drowns out adversarial perturbations while retaining low-frequency components, enhancing robustness without degrading performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the size, number of layers, and number of parameters of the neural network are increased to make it harder for attackers, then security is improved, but the network becomes unnecessarily heavy and difficult to learn, degrading performance

Engineering Contradiction:
ImprovesecurityVSAvoidnetwork complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

A noise network is introduced as an intermediary component between the input and the main neural network. This noise network generates parasitic noise that acts as a mediator to counteract adversarial perturbations, providing security without requiring the main network to be more complex

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If the number of inputs that can be submitted to the neural network is limited or suspicious sequences are detected, then security is improved, but this approach is not always applicable since attackers can legally gain access to the neural network

Engineering Contradiction:
ImprovesecurityVSAvoidaccess flexibility
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The noise network performs preliminary anti-action by pre-processing inputs with parasitic noise before they reach the main neural network. This proactive defense neutralizes adversarial perturbations in advance, allowing the network to maintain both security and unrestricted access to legitimate inputs

Inventive Principle:
Principle #9Preliminary anti-action

Data Source

PatentEP4242922B1Method for securely using a first neural network on an input data
Publication Date: 2025.09.03 IDEMIA PUBLIC SECURITY FRANCE
  • EP4242922B1 patent drawingFigure 1~2

AI summary

The present invention relates to a method for the secure use of a first neural network on an input data, the method being characterized in that it comprises the implementation by data processing means (21) of a terminal (2) of steps of: (a) construction of a second neural network corresponding to the first neural network in which is inserted at least one autoencoder-type neural network trained to add parasitic noise to its input; (b) use of the second neural network on said input data.