Adversarial Corruption for ML Explainability Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing corruption functions in machine learning explainability (MLX) generate only specific types of anomalies, leading to incomplete performance assessments of attribution-based explanation methods, as they fail to produce anomalous tuples with features having small reconstruction errors and large errors in other features, resulting in inaccurate explanations and incomplete testing of explainers.
Innovation Solution
A new corruption function is introduced that modifies a non-anomalous tuple to generate an anomalous tuple with a perturbed feature, causing a decrease in reconstruction error for the perturbed feature and an increase in reconstruction errors for all other features, allowing for the evaluation of explainers' ability to identify the correct causal feature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing corruption functions are used to generate anomalous tuples, then the evaluation process is simple, but the performance assessment of attribution-based explanation methods is incomplete
Solution Approach 1:
The patent changes the parameters of the corruption function by introducing a new approach that modifies multiple features simultaneously rather than single feature corruption. This involves changing the corruption strategy from simple individual feature perturbation to coordinated multi-feature perturbation, thereby improving the comprehensiveness of explainer evaluation while managing complexity through systematic parameter modification.
Solution Approach 2:
The patent segments the evaluation process into multiple stages: first generating base anomalous tuples using existing corruption functions, then applying additional targeted corruptions to create adversarial examples. This segmentation allows the system to build upon simple corruption methods while progressively increasing evaluation rigor, balancing measurement precision with manageable complexity increments.
2Adaptability or versatility
If simple corruption functions are used, then the generation of anomalous tuples is easy, but the reconstruction error distribution is limited
Solution Approach 1:
The patent creates a universal corruption framework that can generate multiple types of anomalies through a single multi-functional corruption function. This function can corrupt individual features, multiple features, or combinations thereof, allowing the same corruption mechanism to produce diverse anomaly types including those with small reconstruction errors on perturbed features and large errors on other features, thereby improving adaptability without requiring separate corruption functions for each anomaly type.
Solution Approach 2:
The patent introduces an intermediary corruption function that acts as a mediator between simple corruption methods and complex evaluation requirements. This intermediary layer applies additional targeted corruptions to tuples generated by simple corruption functions, transforming them into adversarial examples that challenge explainers more rigorously while building upon the simplicity of the base corruption approach.
3Measurement precision
If adversarial inputs are generated to stress-test explainers, then the evaluation accuracy improves, but the complexity of the corruption process increases
Solution Approach 1:
The patent applies preliminary action by first generating anomalous tuples using simple corruption functions, then applying additional targeted corruptions to these pre-generated tuples. This two-stage approach allows the system to prepare base test cases with simple methods, then enhance them with adversarial corruptions, thereby improving evaluation accuracy while managing complexity through staged processing rather than attempting to generate all adversarial examples in a single complex step.
Solution Approach 2:
The patent incorporates feedback mechanisms where the evaluation process identifies weaknesses in explainer performance, and this information feeds back into the corruption function to generate more targeted adversarial examples. The system uses performance metrics from initial evaluations to guide subsequent corruption strategies, improving measurement precision while managing complexity through iterative refinement based on observed explainer behaviors.
Data Source
AI summary
Herein are machine learning (ML) explainability (MLX) techniques that perturb a non-anomalous tuple to generate an anomalous tuple as adversarial input to any explainer that is based on feature attribution. In an embodiment, a computer generates, from a non-anomalous tuple, an anomalous tuple that contains a perturbed value of a perturbed feature. In the anomalous tuple, the perturbed value of the perturbed feature is modified to cause a change in reconstruction error for the anomalous tuple. The change in reconstruction error includes a decrease in reconstruction error of the perturbed feature and/or an increase in a sum of reconstruction error of all features that are not the perturbed feature. After modifying the perturbed value, an attribution-based explainer automatically generates an explanation that identifies an identified feature as a cause of the anomalous tuple being anomalous. Whether the identified feature of the explanation is or is not the perturbed feature is detected.