Outlier Detection in Machine Learning Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems are vulnerable to adversarial attacks, which can produce inaccurate and unexpected results due to non-deterministic execution behavior and black box non-explainability, leading to security concerns in applications like autonomous vehicles and data traffic classification.
Innovation Solution
A method to detect outlier inputs by determining an entropy score derived from classification and activation values, comparing them to a threshold entropy-based distance metric, and adjusting classifications accordingly to flag or block improper inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning systems are used for classification tasks, then productivity and automation are improved, but reliability and security deteriorate due to vulnerability to adversarial attacks
Solution Approach 1:
The patent introduces an intermediary component (outlier detector) between the machine learning classifier and the final decision-making process. This detector analyzes activation values and computes entropy scores to identify adversarial inputs before they can cause harmful misclassifications, thus maintaining security while preserving the automation benefits of ML systems
Solution Approach 2:
The system implements feedback by using the entropy score computation based on activation values to detect outliers and adjust classifications. The feedback loop allows the system to identify when inputs deviate from expected patterns and correct or flag these anomalies, improving reliability without reducing productivity
2Productivity
If machine learning models are trained to distinguish normal traffic patterns, then productivity is improved, but adaptability deteriorates making it difficult to detect unexpected adversarial inputs
Solution Approach 1:
The patent adds another dimension to the classification process by computing entropy scores from activation values across multiple neural network layers. This dimensional extension allows the system to detect adversarial inputs that may appear normal in standard classification but exhibit unusual patterns in the activation space, thus improving adaptability without sacrificing classification speed
3Reliability
If entropy-based outlier detection is implemented, then reliability is improved by detecting adversarial inputs, but device complexity increases
Solution Approach 1:
The system uses self-service by leveraging the existing activation values produced during normal ML classification to compute entropy scores for outlier detection. Rather than requiring separate detection mechanisms or additional training data, the system repurposes the information already generated during classification, thus improving security without proportionally increasing complexity
Data Source
AI summary
A method determines outlier inputs for a machine learning system. The method includes receiving a classification and activation values of a trained classifier or a first input processed by the trained classifier, determining whether an entropy score derived from the first input is below a threshold entropy-based distance metric, and changing the classification in response to the entropy score not being below the threshold.