Federated CNN Parameter Aggregation for Secure Biometric Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current biometric classification methods using convolutional neural networks (CNNs) face challenges in secure data sharing and collaboration between entities, as existing solutions either leak national biometric characteristics or require central manipulation, leading to potential security breaches.

Innovation Solution

A method involving a main server and secondary servers to learn and aggregate CNN parameters from confidential databases while maintaining data confidentiality, using a 'draw and discard' mechanism with noise introduction to create a common representation space for classification, ensuring secure data processing and aggregation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If entities share confidential biometric databases to train CNNs for joint recognition, then recognition effectiveness is improved, but data confidentiality is compromised and security breaches occur

Engineering Contradiction:
Improverecognition effectivenessVSAvoiddata confidentiality
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces a trusted third-party server as an intermediary that coordinates the federated learning process. This server facilitates collaboration between entities by managing model aggregation and training coordination without directly accessing any entity's confidential biometric databases, thus enabling improved recognition effectiveness while preserving data confidentiality through a neutral mediation mechanism

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the centralized training process into distributed local training sessions at each entity's premises. Each entity trains CNN models locally on its own confidential database without sharing the raw data, and only model parameters or gradients are shared with the trusted server for aggregation. This segmentation allows collaboration for improved recognition while maintaining strict data confidentiality at each location

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If a centralized teacher system is used to train student CNNs from multiple confidential databases, then learning from diverse data is improved, but central security vulnerabilities increase and national biometric characteristics may leak

Engineering Contradiction:
Improvelearning from diverse dataVSAvoidcentral manipulation security
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent inverts the traditional teacher-student architecture where a centralized teacher learns from multiple databases. Instead, each entity's local CNN acts as a teacher that trains student CNNs locally on its own confidential data. The student models are then aggregated to create a robust recognition system. This inversion eliminates the need for central manipulation of diverse confidential databases while still enabling learning from diverse data sources through distributed model aggregation

Inventive Principle:
Principle #13The other way round (Inversion)

3Productivity

If CNN parameters are transmitted between servers for joint training, then collaborative recognition is improved, but data transmission security requirements increase

Engineering Contradiction:
Improvecollaborative recognitionVSAvoidtransmission security risks
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent extracts and removes the confidential biometric data from the transmission process entirely. Instead of transmitting raw biometric databases or sensitive information between servers, the system only transmits non-sensitive model parameters, gradients, or aggregation results. This extraction of sensitive data from the transmission channel enables collaborative recognition while dramatically reducing transmission security risks

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3736744B1Method for secure classification of input data by means of a convolutional neural network
Publication Date: 2024.11.20 IDEMIA PUBLIC SECURITY FRANCE
  • EP3736744B1 patent drawingFigure 1
  • EP3736744B1 patent drawingFigure 2
  • EP3736744B1 patent drawingFigure 3

AI summary

The present invention relates to a method for learning parameters of a convolutional neural network (CNN) for data classification, the method comprising implementing, by data processing means (11c) of servers (1a, 1b, 1c), the following steps: (a1) Obtaining the parameters of a set of at least one first CNN; (a2) For a first CNN of said set: - Learning, from a public training database already classified, the parameters of a final representation block (B) of a second CNN corresponding to the first selected CNN to which said representation block (B) has been added; - Retraining, from a confidential training database already classified on a secondary server (1a, 1b), the parameters of the second CNN; - Transmitting to the main server (1c) the parameters of a third CNN corresponding to the second CNN without a final representation block (B);(a3) Replacing a first CNN of said set of first CNNs with the third CNN; (a4) Aggregating said set of at least one first CNN into a fourth CNN. The present invention also relates to a method for classifying input data.