Neural Network Privacy Auditing via Black-Box Loss Binning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks lack effective methods for auditing privacy preservation, particularly in scenarios where the model's privacy guarantees are not directly observable, such as in federated learning or when the model's training data is not shared due to confidentiality concerns.

Innovation Solution

A method and apparatus for auditing neural network privacy by randomly selecting test and training data, calculating a score function difference, binning the results, and estimating privacy levels using total variation distance, allowing for blackbox analysis without requiring knowledge of the neural network's internals.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If differential privacy is implemented by adding random noise to training data, then privacy protection is improved, but the accuracy of the neural network model deteriorates

Engineering Contradiction:
Improveprivacy protectionVSAvoidmodel accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by adding random noise to training data before the neural network training process begins. This pre-processing step ensures that privacy protection is established in advance, and the model is trained on the perturbed data rather than on clean data, thereby maintaining the privacy guarantee while training the model to achieve acceptable accuracy on the noisy data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes the parameter of the training data by introducing random noise with controlled magnitude. By adjusting the noise level (a key parameter), the system can control the trade-off between privacy protection and model accuracy. The noise is added to transform the original training data into perturbed data, fundamentally changing the data parameter while maintaining statistical properties necessary for model training.

Inventive Principle:
Principle #35Parameter changes

2Object-affected harmful factors

If training data is kept confidential and not shared, then data security is improved, but the ability to audit privacy preservation deteriorates

Engineering Contradiction:
Improvedata securityVSAvoidprivacy audit capability
Core Design Contradiction:
Object-affected harmful factorsVSDifficulty of detecting and measuring

Solution Approach 1:

The patent extracts the essential information needed for privacy auditing from the neural network's behavior and output, rather than requiring access to the training data itself. By analyzing the model's performance on test data and comparing it against privacy guarantees, the system can extract auditability information without compromising data security or requiring data sharing.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces an intermediary approach where the neural network model itself serves as the medium for privacy auditing. Instead of directly examining the training data, the audit process uses the model's outputs and behavior patterns as intermediaries to infer privacy preservation characteristics, allowing auditing without data access.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Device complexity

If the neural network operates as a black box without internal access, then system simplicity is improved, but the ability to verify privacy guarantees deteriorates

Engineering Contradiction:
Improvesystem simplicityVSAvoidprivacy guarantee verification
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent applies self-service by enabling the neural network to essentially audit itself through its own behavior patterns. The model's own outputs and performance characteristics provide the necessary information for privacy verification, eliminating the need for external inspection of internal structures while maintaining verification capability.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical approach of directly inspecting internal model structures with a statistical and behavioral approach. Instead of examining the black box internally, the system uses statistical analysis of input-output relationships and performance metrics to verify privacy guarantees, substituting physical inspection with statistical inference.

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

Data Source

PatentEP4654060A1Privacy audit of a neural network
Publication Date: 2025.11.26 NOKIA SOLUTIONS & NETWORKS OY
  • EP4654060A1 patent drawingFigure 1
  • EP4654060A1 patent drawingFigure 2
  • EP4654060A1 patent drawingFigure 3

AI summary

Selecting, at random, test data from a test dataset and training data from a training dataset to obtain an input dataset; inputting the input data into a neural network; determining, for each of the input data, a value of a score function, wherein the score function is an arbitrary loss function; sorting the values of the score function into plural bins; for each of the bins: calculating a proportion of the training data among the input data for which the values of the score function fall in the respective bin; calculating a proportion of the test data among the input data for which the values of the score function fall in the respective; subtracting the proportion of the test data from the proportion of the training data to obtain a distance for the respective bin; calculating a total distance by summing the distances for the bins being larger than 0.