Ensemble ML Models Mitigate Data Poisoning Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems are vulnerable to 'poisoning' attacks where abnormal data is injected, compromising the integrity of training data and decision-making processes, and conventional approaches struggle to accurately detect malfeasance due to limited data and binary labeling methods.
Innovation Solution
A collaborative artificial intelligence system utilizing a population of machine learning models to monitor interactions, generate confidence scores for abnormal data detection, and adaptively reconfigure the model population to remove abnormal data, employing ensemble techniques and retraining to enhance detection accuracy and resilience.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single powerful machine learning model is used for decisioning, then decision-making efficiency is improved, but the system becomes vulnerable to poisoning attacks and abnormal data injection
Solution Approach 1:
The patent divides the single machine learning model into a population of multiple machine learning models. Each model in the population independently analyzes interaction data and generates confidence scores. This segmentation allows the system to maintain decision-making efficiency while improving reliability through diversified analysis, as abnormal data is less likely to consistently deceive all models in the population simultaneously.
2Ease of operation
If conventional binary labeling methods are used for data analysis, then detection simplicity is improved, but detection accuracy of malfeasance deteriorates
Solution Approach 1:
The patent transforms the binary labeling approach into a multi-parameter confidence scoring system. Instead of simply labeling data as normal or abnormal, each machine learning model generates a confidence score indicating the likelihood of abnormal data injection. This parameter change enables more nuanced detection accuracy while maintaining operational simplicity through automated score aggregation and threshold-based decisioning.
3Use of energy by moving object
If limited data is used for training, then training resource consumption is reduced, but the ability to detect abnormal data injection deteriorates
Solution Approach 1:
The patent merges the analytical capabilities of multiple machine learning models to collectively detect abnormal data injection. By combining the confidence scores from multiple models trained on potentially limited data, the system achieves enhanced detection reliability that exceeds what any single model could accomplish alone, effectively pooling their individual detection strengths to overcome data limitations.
4Measurement precision
If a population of machine learning models is deployed to detect poisoning attacks, then detection accuracy is improved, but system complexity increases
Solution Approach 1:
The patent implements feedback mechanisms where the confidence scores generated by the population of machine learning models are aggregated and used to identify and remove abnormal data from training sets. This feedback loop continuously improves detection accuracy by learning from past detections, while the automated nature of the feedback process prevents complexity from becoming unmanageable. The system self-regulates by using its own output to refine its input data quality.
Data Source
AI summary
Embodiments of the present invention provide a system and methods to mitigate poisoning attacks within machine learning systems. The invention includes an improved data analysis approach to train an ensemble of machine learning models to analyze received data and label the data in a non-binary fashion to indicate likelihood that certain data has been injected abnormally and should not be used for training purposes. The resulting dataset from the ensemble is assessed to determine convergence of model labeling and to detect outlier data labeling among models in the ensemble. Confidence scores for clustered interaction data may be performed on varied sets of training data populations and using a number of models. Output from the various training/model mixes are fed to a machine learning model to compare ensemble accuracy between different model sets and select the most accurate ensemble combination.