logwatch-ai 1.0.0

AI-powered Windows Event Log anomaly detection using Isolation Forest

logwatch-ai

AI-powered Windows Event Log anomaly detection library using Isolation Forest.

This library provides a pre-trained machine learning model that analyzes Windows system logs and detects anomalous behavior in real time. It uses a 25-feature Isolation Forest model trained on real Windows Event Log data.

---

Installation

pip install logwatch-ai

Or install from source:

git clone https://github.com/mouayed-kordi/logwatch-ai.git
cd logwatch-ai
pip install .

---

Quick Start

from logwatch_ai import AnomalyDetector

# Initialize the detector (loads the pre-trained model)
detector = AnomalyDetector()

# Score a single log entry
result = detector.score_log(
    level="ERROR",
    message="FATAL: Database connection lost due to timeout",
    log_type="Database",
    cpu=85.2,
    ram=72.1,
)

print(result)
# Output:
# {
#     "is_anomaly": True,
#     "score": -0.1234,
#     "threshold": -0.09
# }

---

API Reference

AnomalyDetector(artifacts_path=None)

Creates a new detector instance and loads the pre-trained model.

Parameter	Type	Description
artifacts_path	str or None	Path to custom .pkl model files. If None, uses the bundled pre-trained model.

detector.score_log(...)

Scores a single log entry for anomaly detection.

Parameter	Type	Default	Description
level	str	(required)	Log severity: "INFO", "WARNING", "ERROR", "CRITICAL"
message	str	(required)	Raw log message text
log_type	str	"Other"	Category: "Database", "Authentication", "Network", "Security", "System", "Other"
cpu	float	0.0	CPU usage % at the time of the log
ram	float	0.0	RAM usage % at the time of the log
timestamp	datetime	now()	When the log occurred
source_key	str	"default"	Identifier to group logs from the same source

Returns a dict:

{
    "is_anomaly": True,    # True if anomaly detected
    "score": -0.1234,      # Raw Isolation Forest score
    "threshold": -0.09     # Calibrated threshold
}

detector.score_batch(logs)

Scores multiple log entries at once.

logs = [
    {"level": "ERROR", "message": "Connection refused", "log_type": "Network"},
    {"level": "INFO", "message": "Service started", "log_type": "System"},
    {"level": "CRITICAL", "message": "Disk full", "log_type": "System", "cpu": 95.0},
]

results = detector.score_batch(logs)
for r in results:
    print(r["is_anomaly"], r["score"])

---

How It Works

The detector builds a 25-feature vector for each log entry:

• Temporal features: Rolling error counts (30s, 1m, 5m, 10m, 15m, 60m)
• Statistical features: CUSUM drift tracking, error density, short/long ratios
• Hardware features: CPU and RAM velocity (rate of change)
• Cyclical time features: Hour and day encoded as sin/cos
• NLP features: TF-IDF + SVD text encoding of the log message

The Isolation Forest model scores each log. If the score falls below the calibrated threshold (-0.09), the log is flagged as anomalous.

---

License

MIT License