Evaluating Model Resilience to Data Poisoning Attacks: A Comparative Study

Machine learning (ML) has become a cornerstone of critical applications, but its vulnerability to data poisoning attacks threatens system reliability and trustworthiness. Prior studies have begun to investigate the impact of data poisoning and proposed various defense or evaluation methods; however, most efforts remain limited to quantifying performance degradation, with little systematic comparison of internal behaviors across model architectures under attack and insufficient attention to interpretability for revealing model vulnerabilities. To tackle this issue, we build a reproducible evaluation pipeline and emphasize the importance of integrating robustness with interpretability in the design of secure and trustworthy ML systems. To be specific, we propose a unified poisoning evaluation framework that systematically compares traditional ML models, deep neural networks, and large language models under three representative attack strategies including label flipping, random corruption, and adversarial insertion, at escalating severity levels of 30%, 50%, and 75%, and integrate LIME-based explanations to trace the evolution of model reasoning. Experimental results demonstrate that traditional models collapse rapidly under label noise, whereas Bayesian LSTM hybrids and large language models maintain stronger resilience. Further interpretability analysis uncovers attribution failure patterns, such as over-reliance on neutral tokens or misinterpretation of adversarial cues, providing insights beyond accuracy metrics.