As Artificial Intelligence (AI) systems transition from academic curiosities to integral components of society, their deployment in safety-critical domains, such as autonomous driving and medical diagnostics, is rapidly accelerating. In these high-sta...
As Artificial Intelligence (AI) systems transition from academic curiosities to integral components of society, their deployment in safety-critical domains, such as autonomous driving and medical diagnostics, is rapidly accelerating. In these high-stakes environments, where errors can have profound, real-world consequences on human well-being and societal trust, ensuring model reliability and trustworthiness becomes a paramount objective. A cornerstone of this endeavor is adversarial robustness—the ability of a model to maintain correct and stable predictions even when faced with small, maliciously crafted perturbations to its input. While this property is essential for building dependable AI, the most common and foundational defense, Adversarial Training (AT), is far from a perfect solution. This dissertation argues that standard AT exhibits significant latent flaws, including compromises in fairness and robust generalization, which can undermine model safety in unexpected and subtle ways. This research directly confronts these challenges. It advances the field by first deeply investigating these critical flaws and then, in a novel pivot, repurposes the very principles of adversarial attacks as a constructive tool for model correction and enhancement.
This research presents a three-part approach. First, in Chapter 3, the critical robust fairness problem is addressed, where standard adversarial training disproportionately harms the robustness of certain classes, creating dangerous blind spots. This dissertation introduces Distance-Aware Fair Adversarial training (DAFA), a novel methodology that makes the training process sensitive to the underlying inter-class similarity. DAFA dynamically assigns class-specific loss weights and adaptive adversarial margins based on the geometric relationships between classes in the feature space. This mechanism effectively improves worst-class robustness and promotes a more equitable defense without degrading average performance. Second, in Chapter 4, the counter-intuitive robust overfitting phenomenon is investigated. This occurs when models begin to memorize hard-to-learn examples during adversarial training, which paradoxically harms robust generalization on unseen data. This research introduces Difficulty Proportional Label Smoothing (DPLS), an adaptive regularization technique designed to combat this. DPLS mitigates memorization by applying a label smoothing factor directly proportional to the measured difficulty of each training example. This intervention forces the model to learn more generalizable, broadly applicable robust features rather than memorizing noisy specifics. Finally, in Chapter 5, this dissertation pivots from defensive training to constructive correction, addressing the pervasive challenge of object hallucination in Large Vision-Language Models (LVLMs). Adversarial Visual Contrastive Decoding (AVCD) is introduced, a novel, training-free inference methodology. AVCD ingeniously repurposes an adversarial attack not to fool the model, but to generate a hard negative visual input—one that maximally induces hallucination. By contrasting the logit distribution from this adversarial, hallucination-prone state against the original, clean-input logits, AVCD precisely suppresses visually ungrounded tokens and significantly enhances model faithfulness at inference time.
Collectively, this dissertation demonstrates the profound versatility of adversarial principles. By first refining adversarial training with data-aware and adaptive techniques (DAFA, DPLS) to shore up its critical weaknesses, and then repurposing these principles for inference-time correction (AVCD), this work provides a robust toolkit. The contributions herein offer tangible pathways for building fairer, more generalizable, and ultimately more reliable AI systems poised for real-world deployment.