Code for Auditing Visualizations: Transparency Methods Struggle to Detect Anomalous Behavior

Imports

After cloning, move inside this repository and:

• create a conda environment using conda env create -f environment.yml
• activate the environment using conda activate audit_vis
• run pip install .

Data

Then, download results.tar.gz, checkpoints.tar.gz and images.tar.gz folders from our Zenodo archive, untar them and move them inside the repository:

• checkpoints contains the weights of the null and anomalous models
• images contains the images we used to compute the explanations
• results contains precomputed visualizations, anomaly scores, and the final results of the paper

Running the scripts

The most important folders in this repository are:

• visualize/, which computes the model explanations and stores them in results/visualizations/
• analysis/, which uses precomputed explanations to compute the anomaly scores and stores them in results/anomaly_scores/

In each of these folders, there is:

• a detection_script.py script, which runs the jobs relevant for the detection task
• a localization_script.py script, which runs the jobs relevant for the localization task

These 4 scripts use submitit to schedule SLURM jobs on a cluster with GPUs. Before running each script:

• set partition_name to the name of your partition
• for the visualize scripts, set the list_anomalies, list_methods
• for the analysis scripts, set the list_anomalies, list_lpips_nets and list_methods variables

Finally, analysis/get_results.py computes the final results for the detection and localization tasks from the anomaly scores.