Reproducibility study of "Robust Fair Clustering: A Novel Fairness Attack and Defense Framework"

TMLR Paper2231 Authors

Abstract: This reproducibility study examines "Robust Fair Clustering: A Novel Fairness Attack and Defense Framework" by Chhabra et al. (2023), an innovative work in fair clustering algorithms. Our study focuses on validating the original paper's claims concerning the susceptibility of state-of-the-art fair clustering models to adversarial attacks and the efficacy of the proposed Consensus Fair Clustering (CFC) defence mechanism. We employ a similar experimental framework but extend our investigations by using additional datasets. Our findings confirm the original paper's claims, reinforcing the vulnerability of fair clustering models to adversarial attacks and the robustness of the CFC mechanism.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

Reviewer 1Q4j pointed out an error in Figures 1 through 4 where not all metrics where equal when 0% of the samples were poisoned. This error was caused by a mistake in the code for plotting, where we labelled the x-axis from 0 through 30%, but entered the data from 12.5 through 30%. We have fixed this and included the datapoints at 0%.

Reviewer 1Q4j also noticed that the pre-attack metrics were varying for the CFC algorithm, while they should be constant. This mistake was caused by the fact that the codebase from the original paper did not apply random seeds for the CFC algorithm, causing variation when running the pre-attack with different percentages of poisoned data. This is also reflected in the Figure 4 of the original paper (Chhabra et al, 2023). We overlooked this problem, and simply re-used their code. However, we have now fixed it by simply taking the mean of all values for the pre-attack for each metric.

Assigned Action Editor: ~Jonathan_Scarlett1

Submission Number: 2231