DeepBridge Robustness Analysis Report

Model: {{ model_name }}
Generated on: {{ timestamp }}
Robustness Summary
Overall Robustness Score
{{ robustness_score }}
Higher is better (0-1 scale)
Avg. Gaussian Impact
{{ raw_impact }}
Lower is better
Avg. Quantile Impact
{{ quantile_impact }}
Lower is better
Base Score
{{ base_score }}
{{ metric }}

Test Configuration

Parameter Value
Model Type {{ model_type }}
Metric {{ metric }}
Iterations {{ iterations }}
Feature Subset {{ feature_subset_display }}
Gaussian Perturbation
Quantile Perturbation
Feature Importance
Model Metrics
Feature Comparison
Worst-Case Analysis
Performance Boxplot
Detailed Results

Gaussian Noise Perturbation Results

Gaussian Perturbation Analysis

This analysis shows how the model performance changes when Gaussian noise is added to input features. The noise level represents standard deviations of the feature distribution.

Primary Model

Noise Level Score Impact Relative Drop (%)

Quantile-based Perturbation Results

Quantile Perturbation Analysis

This analysis shows how the model performance changes when feature values are replaced with values sampled from different quantiles of the distribution.

Primary Model

Perturbation Level Score Impact Relative Drop (%)

Feature Importance for Robustness

Feature Sensitivity Analysis

This analysis shows which features have the most impact on model performance when perturbed. Features with higher scores have greater impact on model robustness.

Feature Importance Score Relative Impact (%)

Model Performance Metrics

Model Metrics

Primary Model

Metric Value

Feature Subset Comparison

All Features vs Feature Subset Comparison

This analysis compares model performance between using all features and using only the selected feature subset under different perturbation levels.

Gaussian Perturbation

Noise Level All Features Score Feature Subset Score Difference

Quantile Perturbation

Perturbation Level All Features Score Feature Subset Score Difference

Worst-Case Performance Analysis

Worst Case Performance by Perturbation Level

This analysis shows the worst-case model performance at each perturbation level across all test iterations.

Gaussian Perturbation - Worst Case

Noise Level Primary Model Feature Subset Alternative Models

Quantile Perturbation - Worst Case

Perturbation Level Primary Model Feature Subset Alternative Models

Performance Distribution Boxplot

Performance Distribution Across Perturbation Levels

This analysis shows the performance distribution at each perturbation level using boxplots, allowing you to visualize the variability in model performance as perturbation increases.

Gaussian Perturbation
Quantile Perturbation
Primary Model

How to interpret:

  • The boxes show the interquartile range (IQR) of performance scores at each perturbation level
  • The horizontal line inside each box represents the median score
  • Whiskers extend to the most extreme data points within 1.5 times the IQR
  • Outliers are plotted as individual points
  • The first box (perturbation level 0.0) represents the baseline performance without perturbation

Detailed Test Results

Run Details