T-LARA: Leveraging Reasoning LLMs to Enhance the Adversarial Robustness of Tabular Security Classifiers

Abstract

Security classifiers in many critical network and application tasks rely on tabular data, yet evaluating their adversarial robustness is complicated by the unique constraints of the tabular domain. Unlike homogeneous image or text data, tabular features possess distinct data types, bounds, intricate interdependencies, and require a highly specific perturbation cost model. Traditionally, developing these feature-level cost models requires extensive manual effort and domain expertise, while reliance on gradient-based computation limits existing methods to specific, differentiable classifier families. In this thesis, we examine whether reasoning-based Large Language Models (LLMs) can circumvent these limitations to produce a classifier-agnostic method for creating realistic, low-cost adversarial examples that adhere to domain constraints. We propose T-LARA, an agentic framework designed to evaluate the adversarial robustness of tabular classifiers. T-LARA automatically constructs a feature-level cost model via LLM reasoning augmented with web-based knowledge acquisition. Driven by this cost model, a Generator agent crafts constraint-preserving adversarial samples, while a Critic agent provides iterative, reasoning-based feedback to refine their effectiveness and quality. Ultimately, this framework demonstrates the potential of reasoning LLMs to drastically reduce the human supervision and domain expertise traditionally required to audit the security of tabular classification systems.