AI Model Poisoned Data Detection via Preliminary Analysis

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

AI models can become untrustworthy if exposed to poisoned training data, leading to inaccurate inferences and requiring costly re-training, which interrupts inference generation and other services by consuming valuable computing resources.

Innovation Solution

A system that analyzes candidate training data against historical data using cluster analysis or similarity measures to detect poisoned data, preventing its use and thus avoiding unnecessary re-training, thereby conserving computing resources for inference generation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If AI models are continuously re-trained to maintain reliability, then model reliability is improved, but computing resources are consumed and service interruption occurs

Engineering Contradiction:

ImproveAI model reliabilityVSAvoidinference generation efficiency

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The system performs preliminary analysis of candidate training data before re-training the AI model. By detecting poisoned data in advance through comparison with historical training data, the system prevents unnecessary re-training operations, thereby maintaining model reliability while avoiding service interruption and preserving computing resources for inference generation.

Inventive Principle:

Principle #10Preliminary action

2Reliability

If computing resources are allocated to re-training AI models, then model reliability is improved, but resource availability for other services is reduced

Engineering Contradiction:

ImproveAI model reliabilityVSAvoidcomputing resources availability

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

3Productivity

If AI models use poisoned training data, then training efficiency is improved, but inference accuracy deteriorates

Engineering Contradiction:

Improvetraining efficiencyVSAvoidinference accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary analysis of candidate training data before re-training the AI model. By detecting poisoned data in advance through comparison with historical training data, the system prevents poisoned data from being used in training, thereby maintaining both training efficiency and inference accuracy.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system introduces an intermediary analysis mechanism that compares candidate training data with historical training data before the training process. This intermediary step identifies poisoned data through similarity analysis, preventing it from affecting the AI model's inference accuracy while allowing clean data to proceed to training.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20240220790A1System and method for preventing introduction of poisoned training data to artificial intelligence models

Publication Date: 2024.07.04 DELL PROD LP

AI summary

Methods and systems for managing artificial intelligence (AI) models are disclosed. To manage AI models, an instance of an AI model may not be re-trained using training data determined to be too similar to previously used training data. By doing so, malicious attacks intending to shift the AI model in a particular direction using poisoned training data may be prevented. To do so, a candidate training data set may be analyzed prior to performing re-training of an instance of an AI model using the candidate training data set. The analysis may result in a score. If the score exceeds a score threshold, the candidate training data set may be considered to contain poisoned training data. If the score does not exceed the score threshold, the candidate training data set may be accepted as usable to train an instance of the AI model.