Global Optimization Benchmarks Enhanced by Black-Box Adversarial Attacks

A new arXiv paper proposes using black-box adversarial attack (BBAA) tasks as benchmarks for global optimization methods, arguing that existing benchmark suites are outdated and limited. The study demonstrates the efficiency of evolutionary algorithms and other metaheuristics on example BBAA problems, aiming to align optimization research with modern machine learning challenges. The paper is categorized under Computer Science > Machine Learning and was submitted to arXiv (ID: 2608.13296). It highlights the need for more relevant and scalable benchmarks in high-dimensional spaces, addressing the risk of bias in method development. The authors advocate for a convergence between global optimization and machine learning, suggesting that BBAA tasks offer a valuable testing ground. The paper includes references, citations, and tools for further exploration, and is part of arXivLabs, a framework for collaborative projects. The work emphasizes openness, community, excellence, and user data privacy, aligning with arXiv's values.