emei-sc-ai/ai-service-classifier

Model Card: AI Service Classifier

Model description

A CPU-friendly scikit-learn multi-task baseline for classifying AI service identity and a synthetic policy-risk category from network telemetry metadata.

Intended use

• Research and demonstration of network-derived AI service visibility.
• Baseline evaluation for approved telemetry pipelines.
• Educational Hugging Face Space or batch-inference demonstrations.

Out-of-scope use

• Identifying people, customers, or organizations.
• Making automated employment, legal, access-control, or incident-attribution decisions.
• Treating the risk output as a confirmed security finding.
• Training on or inferring from raw packet payloads without separate approval.

Training data

Synthetic rows generated by src/ai_service_classifier/synthetic.py. The generator uses simulated DNS/TLS/application metadata, byte and packet statistics, traffic direction, and service profiles. No customer data is included.

Evaluation

Run:

python scripts/evaluate.py

The command writes models/ai-service-classifier/evaluation.json with per-label reports and confusion matrices. Metrics generated from synthetic data should not be represented as production performance.

Bias, risks, and limitations

The synthetic generator encodes simplified service signatures. Real deployments may use proxies, shared CDNs, encrypted metadata, changing domains, or private gateways. The model may therefore misclassify new, rare, or intentionally obfuscated services. Monitor confidence, unknown rates, drift, and false positives; retain human review for high-impact actions.

Privacy and security

This release is designed around metadata only. Never add packet payloads, credentials, customer identifiers, internal hostnames, or proprietary product details without authorization and data-governance review.

License

Apache-2.0 for the repository code. Confirm the final model and dataset release terms before publishing under a company organization.