Fractal Data Division for Robust ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning approaches face challenges in accuracy due to the reliance on frequently occurring features and simplistic methods for combining results from multiple classifiers, which can lead to inaccuracies, especially when adversarial actions conceal these features.
Innovation Solution
The proposed solution involves fractal division of training data among multiple classifiers, allowing each to be trained on different subsets, and fractal aggregation of their results to generate a unified output, thereby enhancing robustness and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple classifiers are trained with the same training data and their results are combined using majority voting, then the system can leverage diverse features from different classifiers, but the accuracy deteriorates when frequently occurring features are concealed by adversaries
Solution Approach 1:
The training data is segmented into multiple disjoint subsets, with each classifier trained on a different subset. This segmentation ensures that classifiers learn from diverse data distributions rather than all training data, making them less susceptible to adversarial concealment of frequently occurring features while maintaining feature diversity across the ensemble.
2Reliability
If classifiers are trained on different subsets of training data, then the system can improve robustness against adversarial features, but the complexity of data management and result aggregation increases
Solution Approach 1:
The aggregation process uses a nested structure where classifiers are organized in levels, with results from lower-level classifiers being aggregated to form inputs for higher-level classifiers. This nested arrangement systematically manages the complexity of combining results from multiple classifiers trained on different data subsets, providing a hierarchical framework for robust aggregation.
3Ease of manufacture
If simple majority voting is used to combine classifier results, then the system is easy to implement, but the accuracy deteriorates when the majority of classifiers are misled by adversarial features
Solution Approach 1:
The system performs preliminary aggregation of classifier results at multiple levels before reaching the final decision. This multi-level preliminary aggregation process allows the system to progressively refine results and mitigate the impact of adversarial features before the final classification, improving accuracy while maintaining implementation feasibility through systematic organization.
4Loss of information
If all classifiers are trained on the complete training dataset, then maximum information is available for training, but the system becomes vulnerable to adversarial features that appear frequently in the training data
Solution Approach 1:
The complete training dataset is segmented into multiple disjoint subsets that are distributed across different classifiers. This segmentation prevents any single classifier from being exposed to all training examples, including adversarial features, while collectively the ensemble still utilizes information from the complete dataset through their diverse training subsets.
Data Source
AI summary
Systems, computer-implemented methods, and computer program products to aggregate classifier result are provided. According to an embodiment, a system can comprise a memory that stores computer executable components and a processor that executes the computer executable components stored in the memory. The computer executable components can comprise a fractal dividing component that can train classifiers with fractally defined disjoint and non-disjoint subsets of training data comprising data with a first label and data with a second label. Further, the system can include training component that can train the group of classifiers with the fractally defined disjoint and non-disjoint subsets of training data. A fractal aggregating component can be included to generate a unified result of the classifiers by fractally aggregating results of classification by the ones of the classifiers, based on the subsets.