Multi-Label Black-Box Attacks via Evolutionary Structured Many-Objective Adversarial Perturbations

Multi-label learning poses significant challenges due to the complexities of co-occurring labels. Adversarial examples are critical in safety-sensitive domains, where malicious tampered data can compromise models. Yet, their application to tabular multi-label learning remains under-explored, presenting a potential security risk. This paper introduces an adversarial training framework leveraging Evolutionary Computation, specifically the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), to craft structured and concealable adversarial examples for tabular multi-label classifiers. Our contributions centre on an effective adversarial method tailored for tabulated multi-label learning, and a many-objective framework to balance the conflict between multi-label attack success, attack robustness, and attack concealability. We extend multi-label adversarial training to cope with tabulated data necessitating novel methods for generating structured adversarial examples and assessing attack concealability compared to image-based approaches—pioneering future research in tabulated multi-label adversarial training.Our framework also simulates real-world black-box attack scenarios where true model information is unknown. Our approach trains adversarial examples without prior knowledge of the target model by competing with a proxy model, progressively training more robust adversarial examples. Experiments show a high attack success rate (81.3–100%) across large datasets, significantly reducing multi-label classification performance post-perturbation and confirming the concealability of the attacks. Our results highlight the robustness of our approach, advancing adversarial training for multi-label, tabulated data.