waveguard-benchmarks
收藏资源简介:
WaveGuard异常检测基准数据集是一个用于评估异常检测模型的精选基准数据集。该数据集包含标记的训练集(正常样本)和测试集(正常与异常混合样本),以及WaveGuard与传统方法的对比结果。数据集涵盖12个不同的场景,包括信用卡欺诈、网络入侵、加密货币欺诈等。具体包含两个子数据集:1) 服务器指标数据集,模拟服务器健康指标并注入故障事件,包含CPU、内存、磁盘I/O、网络和错误等5个数值特征;2) 合成时间序列数据集,包含正弦波、趋势、季节性和随机游走等模式,并注入点异常、上下文异常和集体异常。数据集以Parquet格式提供,每个子数据集包含训练集、测试集和元数据文件。适用于表格分类、时间序列分类、异常检测等任务,尤其适用于服务器监控和网络安全领域。
The WaveGuard Anomaly Detection Benchmark Dataset is a curated benchmark dataset for evaluating anomaly detection models. This dataset includes labeled training sets (normal samples) and test sets (mixed samples of normal and anomalous instances), as well as comparative results between WaveGuard and traditional anomaly detection methods. The dataset covers 12 distinct scenarios, including credit card fraud, network intrusion, cryptocurrency fraud, and others. Specifically, it contains two sub-datasets: 1) Server Metrics Dataset, which simulates server health metrics and injects fault events, with 5 numerical features including CPU, memory, disk I/O, network traffic, and error logs; 2) Synthetic Time Series Dataset, which includes patterns such as sine waves, trends, seasonality, and random walks, and injects point anomalies, contextual anomalies, and collective anomalies. The dataset is provided in Parquet format, and each sub-dataset contains training set, test set, and metadata files. It is applicable to tasks such as tabular classification, time series classification, and anomaly detection, and is particularly suitable for server monitoring and cybersecurity fields.
WaveGuard Anomaly Detection Benchmarks 数据集概述
数据集基本信息
- 名称:WaveGuard Anomaly Detection Benchmarks
- 发布者:Partin, Greg
- 发布日期:2026年
- 许可证:MIT
- 数据规模:1K<n<10K
- 任务类别:表格分类
- 标签:异常检测、时间序列、时间序列分类、服务器监控、网络安全、基准测试、waveguard、零训练
- 官方名称:WaveGuard Anomaly Detection Benchmarks
数据集内容与目的
该数据集是一个用于评估异常检测模型的精选基准数据集和对比结果集合。它包含带标签的训练集(正常样本)和测试集(正常与异常混合样本),以及WaveGuard与传统方法的直接对比结果。
包含的数据集
1. Server Metrics (server_metrics/)
- 描述:模拟的服务器健康指标,注入了故障事件。
- 特征:cpu, memory, disk_io, network, errors (5个数值特征)
- 训练集:500个正常样本
- 测试集:100个样本(其中15个异常)
- 异常类型:CPU峰值、内存泄漏、磁盘饱和、网络洪泛
2. Synthetic Time Series (synthetic_timeseries/)
- 描述:具有已知异常注入点的受控合成信号。
- 模式:正弦波、趋势、季节性、随机游走
- 异常类型:点异常(峰值)、上下文异常(细微偏移)、集体异常(状态变化)
- 训练集:每种模式200个干净窗口
- 测试集:每种模式50个窗口(每种10个异常)
数据格式
每个数据集以Parquet文件形式提供:
dataset_name/ train.parquet # 仅正常样本 test.parquet # 正常与异常混合样本 metadata.json # 特征描述、异常计数、创建参数
基准测试对比结果
WaveGuard与IsolationForest、LOF和OneClassSVM在12个数据集上的对比。 总结:在12个数据集中,WaveGuard在11个数据集的F1分数上排名第一。
| 数据集 | WaveGuard | IsolationForest | LOF | OneClassSVM | 优胜者 |
|---|---|---|---|---|---|
| Credit Card Fraud* | 0.653 | 0.607 | 0.601 | 0.472 | WaveGuard |
| Network Intrusion* | 0.598 | 0.252 | 0.232 | 0.546 | WaveGuard-L1 |
| Crypto Fraud | 1.000 | 0.933 | 0.946 | 0.897 | WaveGuard |
| Prompt Injection | 0.976 | 0.952 | 0.976 | 0.889 | WaveGuard |
| Phish Guard | 0.976 | 0.905 | 0.952 | 0.816 | WaveGuard |
| Content Guard | 0.975 | 0.842 | 0.879 | 0.784 | WaveGuard |
| Fraud Lens | 0.949 | 0.896 | 0.882 | 0.800 | WaveGuard |
| Ad Click Fraud | 0.988 | 0.952 | 0.930 | 0.889 | WaveGuard |
| Insurance Claims | 0.972 | 0.921 | 0.959 | 0.833 | WaveGuard |
| Network Security | 0.990 | 0.962 | 0.980 | 0.952 | WaveGuard |
| API Monitoring | 0.959 | 0.909 | 0.933 | 0.814 | WaveGuard |
| Log Anomalies | 0.946 | 0.875 | 0.875 | 0.805 | WaveGuard |
*真实世界数据集。其他数据集使用具有现实特征模式的领域特定测试套件。
详细结果(包括样本大小、特征数量和异常率)请参见 benchmark_results/comparison.json。
使用方式
python from datasets import load_dataset
ds = load_dataset("emergentphysicslab/waveguard-benchmarks", "server_metrics") train = ds["train"].to_pandas() test = ds["test"].to_pandas()
评估协议
- 仅在
train.parquet上训练/拟合检测器 - 对
test.parquet中的每一行进行评分 - 报告:精确率、召回率、F1分数、AUC-ROC、平均延迟
- 与模型卡中的WaveGuard基线进行比较
引用信息
bibtex @dataset{waveguard_benchmarks2026, title={WaveGuard Anomaly Detection Benchmarks}, author={Partin, Greg}, year={2026}, url={https://huggingface.co/datasets/emergentphysicslab/waveguard-benchmarks} }




