Federated Learning Risk Audits for Poisoning-Resistant Model Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing federated learning methods are vulnerable to attacks from malicious client devices that can poison and bias the global AI model, leading to data reconstruction vulnerabilities, and traditional security measures like encrypted communications do not scale well with large numbers of devices, while Trusted Execution Environments (TEEs) face resource limitations and complex communication requirements.

Innovation Solution

Implementing electronic risk audits on client devices to assess their risk levels, applying parameter privacy adjustment methods based on risk scores to configure a global AI model, iteratively refining these adjustments until risk scores are below a threshold, and achieving a balanced tradeoff between utility and risk.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If static noise injection is used to prevent data reconstruction attacks, then security is improved, but model performance deteriorates

Engineering Contradiction:
ImprovesecurityVSAvoidmodel performance
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The patent implements dynamic noise injection where the noise level is adjusted based on the risk score of each client device. High-risk devices receive stronger noise injection while low-risk devices receive minimal or no noise, making the security mechanism adaptive rather than static. This resolves the contradiction by making security measures proportional to actual risk levels.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent applies different noise injection levels to different client devices based on their individual risk assessments. Each device receives a customized level of privacy protection tailored to its risk profile, rather than applying uniform noise injection to all devices. This local differentiation maintains performance for trustworthy devices while securing against malicious ones.

Inventive Principle:
Principle #3Local quality

2Reliability

If traditional encrypted communications are used to secure federated learning, then security is improved, but scalability deteriorates

Engineering Contradiction:
ImprovesecurityVSAvoidscalability
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent extracts the security verification process from the main federated learning communication flow by implementing lightweight risk audits that operate independently. The risk scoring mechanism separates security assessment from model parameter exchange, allowing encrypted communications to scale while maintaining security through asynchronous risk evaluation.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the security approach from heavy cryptographic protocols to a parameter-based risk scoring system. By using numerical risk scores derived from audit metrics rather than full encryption overhead, the system achieves security with minimal impact on communication scalability and system productivity.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If Trusted Execution Environments are implemented to protect model privacy, then security is improved, but device complexity and resource requirements worsen

Engineering Contradiction:
Improvemodel privacyVSAvoidresource requirements
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces expensive, complex TEE infrastructure with lightweight, software-based risk audit mechanisms. The risk scoring system uses computationally inexpensive metrics and statistical methods instead of requiring hardware-based trusted execution environments, making privacy protection accessible to resource-constrained devices.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Solution Approach 2:

The patent introduces a central computing device that acts as an intermediary to perform risk audits and generate risk scores. This mediator handles the complex security verification centrally, allowing client devices to remain simple without TEEs while still achieving model privacy through the intermediary's risk-based noise injection control.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Reliability

If uniform parameter privacy adjustment is applied to all client devices, then security is improved, but model accuracy deteriorates

Engineering Contradiction:
ImprovesecurityVSAvoidmodel accuracy
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The patent applies local quality by differentiating privacy adjustment levels for each client device based on individual risk scores. Low-risk devices contribute parameters with minimal or no noise injection, preserving model accuracy from trustworthy sources, while high-risk devices have stronger privacy adjustments applied to mitigate potential attacks. This localized approach resolves the contradiction between security and accuracy.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent makes the parameter privacy adjustment dynamic by continuously updating risk scores based on ongoing audits and adjusting noise injection levels accordingly. The system adapts the degree of privacy protection for each device over time rather than applying fixed uniform adjustments, allowing model accuracy to be maintained from reliable devices while security is enhanced for problematic ones.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4682769A1Device, system and method for federated learning using risk audits
Publication Date: 2026.01.21 AMADEUS SAS
  • EP4682769A1 patent drawingFigure 1
  • EP4682769A1 patent drawingFigure 2
  • EP4682769A1 patent drawingFigure 3

AI summary

A computing device, that is configured to configure a global machine learning model, performs respective electronic risk audits of client devices configured to train respective local machine learning models that correspond to a global machine learning model. Based on respective electronic risk scores of one or more of the client devices, determined via the respective electronic risk audits, the computing device implements one or more parameter privacy adjustment methods on respective parameters received from the client devices prior to using the respective parameters to configure the global machine learning model, wherein respective client devices determined to have higher electronic risk scores have more of the parameter privacy adjustment methods applied than other respective client devices determined to have lower electronic risk scores. The computing device provides, to the client devices, the global machine learning model configured according to the respective parameters as adjusted.