emergentphysicslab/waveguard-benchmarks

WaveGuard Anomaly Detection Benchmarks

Curated benchmark datasets and comparison results for evaluating anomaly detection models. Includes labeled training (normal) and test (mixed normal + anomalous) splits, plus head-to-head comparisons between WaveGuard and traditional methods.

Benchmark Comparisons (benchmark_results/)

WaveGuard vs. IsolationForest, LOF, and OneClassSVM across 12 datasets.

Summary: WaveGuard ranked #1 on all 12 datasets by F1 score.

Dataset	WaveGuard	IsolationForest	LOF	OneClassSVM	Winner
Credit Card Fraud*	0.653	0.607	0.601	0.472	WaveGuard
Network Intrusion*	0.598	0.252	0.232	0.546	WaveGuard
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

*Real-world datasets. Others use domain-specific test suites with realistic feature schemas.

See benchmark_results/comparison.json for full details including sample sizes, feature counts, and anomaly rates.

Datasets

1. Server Metrics (server_metrics/)

Simulated server health metrics with injected failure events.

• Features: cpu, memory, disk_io, network, errors (5 numeric)
• Training: 500 normal samples
• Test: 100 samples (15 anomalous)
• Anomaly types: CPU spike, memory leak, disk saturation, network flood

2. Synthetic Time Series (synthetic_timeseries/)

Controlled synthetic signals with known anomaly injection points.

• Patterns: sinusoidal, trend, seasonal, random walk
• Anomaly types: point (spike), contextual (subtle shift), collective (regime change)
• Training: 200 clean windows per pattern
• Test: 50 windows per pattern (10 anomalous each)

Format

Each dataset is provided as Parquet files:

dataset_name/
  train.parquet     # Normal samples only
  test.parquet      # Mixed normal + anomalous
  metadata.json     # Feature descriptions, anomaly counts, creation params

Usage

from datasets import load_dataset

ds = load_dataset("emergentphysicslab/waveguard-benchmarks", "server_metrics")
train = ds["train"].to_pandas()
test = ds["test"].to_pandas()

Evaluation Protocol

1. Train/fit your detector on train.parquet only
2. Score each row in test.parquet
3. Report: Precision, Recall, F1, AUC-ROC, Average Latency
4. Compare against WaveGuard baseline in the model card

Citation

@dataset{waveguard_benchmarks2026,
  title={WaveGuard Anomaly Detection Benchmarks},
  author={Partin, Greg},
  year={2026},
  url={https://huggingface.co/datasets/emergentphysicslab/waveguard-benchmarks}
}