CARBEN

CARBEN是由国立清华大学开发的复合对抗性鲁棒性基准，旨在评估和提升模型在复合扰动下的鲁棒性。该数据集通过集成多种威胁模型，如颜色、亮度、对比度等，模拟真实世界中的复杂攻击场景。创建过程中，CARBEN允许用户交互式地调整攻击参数，实时观察模型预测的变化。应用领域主要集中在提高深度神经网络在面对复杂对抗攻击时的鲁棒性，解决模型在实际应用中的安全问题。

CARBEN is a comprehensive adversarial robustness benchmark developed by National Tsing Hua University, aimed at evaluating and enhancing the robustness of models under comprehensive perturbations. This dataset simulates complex real-world attack scenarios by integrating multiple threat models such as color, brightness, contrast and others. During its development, CARBEN allows users to interactively adjust attack parameters and observe changes in model predictions in real time. Its application scenarios mainly focus on enhancing the robustness of deep neural networks against complex adversarial attacks, and addressing the security issues of models in practical applications.