Quantized CNN Parameter Security via Segmented Whitening

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Convolutional Neural Networks (CNNs) deployed on resource-constrained devices like mobile terminals face security vulnerabilities due to the large size of required tables for whitened implementations, making it challenging to secure the storage and manipulation of sensitive parameters, especially against side-channel and fault attacks.

Innovation Solution

A method for learning quantized, fully convolutional CNN parameters with separable convolutions and binary representations, which reduces the size of parameters and enables whitened implementation on standard hardware by using a server for learning and a terminal for classification, employing a whitened implementation of convolution, pooling, and nonlinear layers to secure parameter storage and manipulation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If whitened implementation is used to secure CNN parameters, then security against side-channel and fault attacks is improved, but the table size becomes colossal and unsuitable for resource-constrained devices

Engineering Contradiction:
ImprovesecurityVSAvoidtable size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies segmentation by dividing the CNN into separate functional layers (convolutional layers, pooling layers, nonlinear layers) and applying whitened implementation to each layer independently. This segmentation allows the overall secure system to be built from smaller, manageable secure components, reducing the memory burden while maintaining security.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes parameters by quantizing CNN parameters to discrete values and representing them in binary form. This parameter transformation reduces the precision requirements while maintaining functional equivalence, thereby reducing the size of tables needed for whitened implementation without significantly degrading classification performance.

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If quantized CNN with separable convolutions is used, then parameter size is reduced for resource-constrained devices, but classification precision may be degraded

Engineering Contradiction:
Improveparameter sizeVSAvoidclassification precision
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

Separable convolutions segment the convolution operation into two independent steps: depthwise convolution (spatial filtering) and pointwise convolution (channel mixing). This segmentation reduces the number of parameters from O(k^2 * C_in * C_out) to O(k^2 * C_in + C_in * C_out), achieving compression while maintaining classification precision through the two-stage processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies quantization by transforming continuous CNN parameters into discrete values with limited precision (e.g., 8-bit or lower). This parameter change reduces memory storage requirements and computational complexity while using techniques like binarization and quantized arithmetic to maintain classification accuracy within acceptable ranges.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3511870B1Methods for learning of parameters of a convolutional neural network, and classification of input data
Publication Date: 2024.09.04 IDEMIA PUBLIC SECURITY FRANCE
  • EP3511870B1 patent drawingFigure 1
  • EP3511870B1 patent drawingFigure 2a
  • EP3511870B1 patent drawingFigure 2b

AI summary

The present invention relates to a method for learning parameters of a convolutional neural network (CNN) for data classification. The method comprises implementing, by means of data processing equipment (11) of a server (1), the steps of: (a1) Learning, from a training database already classified, the parameters of a quantized CNN, such that said parameters have values ​​in a discrete space; (a2) Generating a whitened implementation of at least one layer of said quantized CNN, said whitened implementation being predetermined based on at least one of said learned parameters. The present invention also relates to a method for classifying input data.