Towards Responsible Machine Learning: Enhancing Robustness Against Adversaries and Corruptions and Promoting Fairness in Decision-Making
Year: 2024
Type: dissertation
Abstract: Designing trustworthy machine learning (ML) systems requires addressing critical challenges in robustness and fairness. Robust ML models must withstand corruptions, adversarial perturbations and other uncertainties to ensure reliable performance, while fairness mechanisms aim to mitigate biases and promote equitable outcomes across diverse scenarios. This thesis explores these two distinct but vital dimensions of responsible ML, presenting novel approaches to enhance robustness in ML models and long-term fairness in decision-making systems.We first investigate fundamental defenses against adversarial attacks by studying a simplified Gaussian hypothesis testing scenario. In this setting, adversarial perturbations are modeled as nuisance parameters, and we propose a defense using the generalized likelihood ratio test to jointly estimate the class of interest and the perturbation. This approach achieves asymptotic performance comparable to minimax defenses while offering a favorable robustness-accuracy trade-off in non-asymptotic regimes. Next, building on insights from communication theory, we introduce a novel approach to combating out-of-distribution corruptions and adversarial perturbations in deep neural networks. Our tilted exponential (TEXP) layer and training objective enhance robustness by promoting sparse and meaningful layer-wise representations, improving the signal-to-noise ratio during learning. We show through experiments on standard image datasets that TEXP learning integrates with standard empirical risk minimization and achieves robustness without relying on data augmentation. In a complementary study, we address fairness in sequential decision-making processes, such as hiring and college admissions, where fairness must evolve and persist over time. Unlike one-shot fairness metrics, sequential decision-making introduces feedback loops where biases in early decisions can influence future applicant pools. We examine the hypothesis that slightly favoring underrepresented groups in selection can generate positive reinforcement, increasing diversity in subsequent rounds. We propose the Multi-agent Fair-Greedy policy, which balances fairness with score maximization, and prove its convergence to fairness targets under certain conditions. Empirical results on synthetic and adapted real-world datasets demonstrate its effectiveness, but we also highlight the risks of uncoordinated agent behavior in more complex population dynamics, where negative feedback loops may reduce fairness. These findings emphasize the need for policy makers to design carefully to respond to evolving social and demographic conditions.Although robustness and fairness address distinct challenges in responsible ML, both are essential for creating reliable and socially responsible systems. This work provides foundational insights and practical tools to tackle these dimensions, advancing the development of ML systems for real-world applications.
Author: Puranik, Bhagyashree
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6987609456