Kyoto-2006+

Kyoto-2006+是一个用于网络入侵检测的大规模数据集，由Bitdefender, Romania创建。该数据集覆盖了10年的真实网络流量数据，包含自然发生的随时间变化的数据，如用户行为模式的变化和软件更新。数据集的创建旨在研究无监督异常检测中的分布偏移问题，通过分析非平稳数据特性，使用t-SNE和最优传输方法测量不同年份之间的分布距离。AnoShift基准通过将数据分为IID、NEAR和FAR测试分割，验证了模型在时间上的性能退化，并展示了通过承认和处理分布偏移问题，性能可以得到改善，平均提升可达3%。

Kyoto-2006+ is a large-scale dataset for network intrusion detection, created by Bitdefender of Romania. This dataset contains 10 years of real-world network traffic data, including naturally occurring time-varying phenomena such as shifts in user behavior patterns and software updates. The dataset was developed to study the distribution shift problem in unsupervised anomaly detection: t-SNE and optimal transport methods are used to measure distribution distances across different years by analyzing the non-stationary characteristics of the data. The AnoShift benchmark splits the dataset into IID, NEAR and FAR test splits to validate model performance degradation over time, and demonstrates that acknowledging and addressing distribution shift issues can improve model performance, with an average improvement of up to 3%.