Auditing Visualizations: Transparency Methods Struggle to Detect Anomalous Behavior

Data from the submission Auditing Visualizations: Transparency Methods Struggle to Detect Anomalous Behavior:

• 'checkpoints' contains the weights of the models we analyzed: 12 null models and 24 anomalous models
• 'images' contains the images we obtained explanations for: 
  • The 50 images sampled from ImageNet for the detection task
  • The normal and abnormal images we used for each anomalous model for the localization task
• 'results' contains the results of our analysis:
  • The visualizations of all images by the different explanations methods
  • Anomaly scores for each method, image, and model
  • Overall results for the detection and localization tasks