Efficient Estimators for Average-Case Robustness

Code repository for the paper "Characterizing Data Point Vulnerability via Average-Case Robustness (UAI 2024) by Tessa Han*, Suraj Srinivas*, and Hima Lakkaraju".

Summary

Average-case robustness ($p^\mathrm{robust}_{\sigma}$) is a useful measure of data point vulnerability. However, the naïve approach to compute $p^\mathrm{robust}_{\sigma}$ (i.e., $p^\mathrm{mc}_{\sigma}$), which relies on Monte Carlo sampling, is computationally intractable. Thus, in this paper, we develop efficient analytical estimators of average-case robustness ($p^\mathrm{taylor}_{\sigma}, p^\mathrm{mmse}_{\sigma}, p^\mathrm{taylor_mvs}_{\sigma}$ and $p^\mathrm{mmse_mvs}_{\sigma}$). We demonstrate the accuracy and efficiency of these analytical estimators for standard deep learning models and their usefulness for identifying vulnerable data points and quantifying robustness bias of models.

Demo

The $p^\mathrm{robust}_{\sigma}$ estimators are implemented in the estimators folder. A demonstration of how to use each estimator can be found in calc_p_robust.py. To run calc_p_robust.py:

• Navigate into the average-case-robustness repo
• Run $ python calc_p_robust.py

This will print the following output (numbers may vary):

--> load data
Files already downloaded and verified
--> load model
--> calculate p_robust
p_mc: [0.78, 0.88, 0.3, 0.36, 1.0, 1.0, 0.98, 1.0]
p_tay: [1.0, 0.99, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
p_tay_mvs: [1.0, 0.92, 0.92, 0.94, 1.0, 0.97, 0.88, 0.94]
p_mmse: [0.96, 0.79, 0.13, 0.12, 1.0, 1.0, 1.0, 1.0]
p_mmse_mvs: [0.84, 0.62, 0.2, 0.18, 1.0, 1.0, 0.95, 0.99]
p_softmax: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
complete!

Citation

@inproceedings{averagecaserob2024,
    title={Characterizing Data Point Vulnerability via Average Case Robustness},
    author={Han*, Tessa, and Srinivas*, Suraj, and Lakkaraju, Himabindu},
    booktitle={Conference on Uncertainty in Artificial Intelligence (UAI)},
    year={2024}
}