Keywords: adversarial attacks, fairness
TL;DR: Reproduction of the paper "Exacerbating Algorithmic Bias through Fairness Attacks"
Abstract: Scope of Reproducibility: We conducted a reproducibility study of the paper "Exacerbating Algorithmic Bias through Fairness Attacks". According to the paper, current research on adversarial attacks is primarily focused on targeting model performance, which motivates the need for adversarial attacks on fairness. To that end, the authors propose two novel data poisoning adversarial attacks, the influence attack on fairness and the anchoring attack. We aim to verify the main claims of the paper, namely that: a) the proposed methods indeed affect a model's fairness and outperform existing attacks, b) the anchoring attack hardly affects performance, while impacting fairness, and c) the influence attack on fairness provides a controllable trade-off between performance and fairness degradation. Methodology: We chose PyTorch Lightning to re-implement all of the code required to reproduce the original paper's results. Our implementation enables the quick and easy extension of existing experiments, as well as the integration with the various development tools that come with PyTorch Lightning. All of our experiments took about 120 hours to complete on a machine equipped with an Intel Core i7 7700k CPU and an NVIDIA GeForce GTX 1080 GPU. Results: Our results slightly deviate from the ones reported by the authors. This could be attributed to the design choices we had to make, due to ambiguities present in the original paper. After inspecting the provided codebase along with relevant literature, we were able to replicate the experimental setup. In our experiments, we observe similar trends and hence we can verify most of the paper's claims, albeit not getting identical experimental results. What was easy: The original paper is well-structured and easy to follow, with the principle ideas behind the proposed algorithms being very intuitive. Additionally, the datasets used in the experiments are publicly available, small in size, and the authors provide their code on GitHub. What was difficult: During our study, we encountered a few unforeseen issues. Most importantly, we were not able to identify critical technical information required for the implementation of the proposed algorithms, as well as a detailed description of the models used, their training pipeline, hyperparameters, and data pre-processing techniques. Furthermore, the publicly available code is convoluted and employs out-of-date libraries, making it difficult to set up the necessary environment. Communication with original authors: We contacted the paper's first author once to confirm our understanding of certain elements of the paper that were either not specific enough or missing. Although they responded fairly quickly, their answer prompted us back to the paper and the provided codebase, while not encouraging any further communication.
Paper Url: https://ojs.aaai.org/index.php/AAAI/article/view/17080
Paper Venue: AAAI 2021