ML Model Robustness Assessment Against Extraction Attacks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine Learning (ML) models are vulnerable to model extraction attacks, where attackers can replicate the model's performance using limited query-response access, leading to privacy violations and legal consequences, with existing evaluation methods focusing on performance rather than resilience against such attacks.

Innovation Solution

A method to assess the robustness and resilience of ML models against model extraction attacks by training candidate substitute models with an external dataset, evaluating their performance using various metrics, and calculating the similarity with the original model, simulating a black-box scenario with a query budget constraint to determine the model's robustness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If query-response access is provided to ML models for legitimate use, then model utility and accessibility are improved, but vulnerability to model extraction attacks increases

Engineering Contradiction:
Improvemodel accessibilityVSAvoidmodel security
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent performs robustness assessment before deploying the ML model in production. By training substitute models and evaluating their performance in advance, the system identifies vulnerability to extraction attacks before the model is exposed to attackers, allowing preventive measures to be taken while maintaining query-response access for legitimate use

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements a feedback mechanism where the robustness assessment results are used to inform deployment decisions. The system continuously evaluates whether the model meets security thresholds before making it accessible, creating a closed-loop process that balances accessibility with security

Inventive Principle:
Principle #23Feedback

2Device complexity

If standard performance evaluation methods are used for ML models, then model performance measurement is simplified, but detection of privacy violations and extraction attacks is insufficient

Engineering Contradiction:
Improveevaluation complexityVSAvoidsecurity assessment accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the evaluation process into two distinct parts: standard performance evaluation (accuracy, precision, recall) and security robustness evaluation (extraction attack resistance). This segmentation allows both simple performance measurement and comprehensive security assessment to coexist, with each serving its specific purpose

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary assessment layer that bridges standard performance metrics and security evaluation. By using substitute models as intermediaries to simulate extraction attacks, the system can measure security robustness without directly compromising the original model, while still using standard evaluation frameworks

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If multiple candidate substitute models are trained to assess robustness, then assessment comprehensiveness is improved, but computational resources and time are increased

Engineering Contradiction:
Improveassessment comprehensivenessVSAvoidassessment efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent trains multiple candidate substitute models using different learning algorithms to ensure comprehensive assessment coverage. By evaluating multiple candidates rather than a single model, the system achieves more reliable robustness measurement, accepting the additional computational cost as necessary for thorough security assessment

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent varies parameters in the robustness assessment process, including different query budget constraints, different learning algorithms for substitute models, and different evaluation metrics. This parameter variation enables comprehensive assessment while allowing the system to identify patterns and trends that improve overall assessment efficiency

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4365787A1A method for assessing robustness and resilience of machine learning models to model extraction attacks on ai-based systems
Publication Date: 2024.05.08 DEUTSCHE TELEKOM AG
  • EP4365787A1 patent drawingFigure 1
  • EP4365787A1 patent drawingFigure 2
  • EP4365787A1 patent drawing

AI summary

A system for performing an assessment of the robustness and resilience of an examined original ML model against model extraction attacks, comprising a computerized device having at least one processor, which is adapted to train multiple candidate models MC with the external dataset D for each of the specified candidate learning algorithms α in Alg, where each candidate substitute model is trained on a subset of D corresponding to the evaluated ith query limit of the query budget constraint Q; evaluate the performance of each substitute model MC according to different evaluation methods ∈ Evaluation; calculate the robustness of each substitute model, where smaller difference or high agreement/similarity rate between the performance of the original model and the substitute model indicates that the original and substitute models are similar to each other, and that the substitute model having the highest performance can mimic the behavior of the original model and can be used as a replica of the original model.