PulseNet-RUL-Forecasting

Remaining Useful Life forecasting on NASA C-MAPSS FD001 — built as a secure MLOps reference implementation with a STRIDE threat model, RBAC, tenant audit trail, and SARIF CI security gates.

This is the ops side of ML security: what does a properly hardened ML inference pipeline look like from data ingestion to prediction API?

---

Threat model (STRIDE)

Full STRIDE analysis: STRIDE_THREAT_MODEL.md

STRIDE category	Threat	Mitigation
Spoofing	Sensor spoofing — malicious actors inject false telemetry readings	Mutual TLS + API key auth; data origin validation at ingestion gate
Tampering	Training data or model weights modified at rest	SHA-256 hash chain on pipeline artifacts; Ed25519-signed model checkpoints
Repudiation	Actor denies performing a critical action (model update, data deletion)	Append-only structured audit log — tenant, user, timestamp, prediction, input hash
Information disclosure	Proprietary sensor data or model IP exfiltrated	RBAC per tenant; encryption in transit and at rest; 0.019 ms mean encrypt overhead
Denial of service	Prediction endpoint flooded	Rate limiting; load balancing config in docker-compose.yml
Elevation of privilege	Low-privilege user gains admin access	Least-privilege RBAC enforced at API layer; scoped JWT tokens

---

What this implements

Component	Implementation	Location
RUL model	LSTM + Transformer ensemble on C-MAPSS FD001	src/pulsenet/models/
Data lineage	SHA-256 hash chain from raw sensor data to training set	src/pulsenet/pipeline/
Input validation	Schema enforcement + anomaly gate before inference	src/pulsenet/security/adversarial_telemetry_guard.py
RBAC	Role-scoped API endpoints; tenant isolation at middleware layer	src/pulsenet/api/auth.py, src/pulsenet/api/middleware/tenant.py
Audit trail	Append-only JSONL log — prediction, input hash, tenant, user, timestamp	src/pulsenet/security/audit.py
Encryption	AES-GCM at rest; TLS in transit	src/pulsenet/security/encryption.py
CI gates	GitHub Actions: dependency scan + artifact hash check + SARIF output	.github/workflows/ci.yml
NIST AI RMF controls	Mapped controls in docs/nist_ai_rmf_controls.yaml	docs/

---

Benchmark results

Dataset: NASA C-MAPSS FD001 (official .zip — data/official/CMAPSSData.zip)

Inference latency

Metric	Value
Mean	2.7 ms
Median	2.5 ms
P95	3.9 ms
P99	4.3 ms
Target met	✓

Throughput (samples/sec)

Batch size	Throughput
1	329
32	13,429
128	31,424
256	52,368

Anomaly detection (adversarial telemetry guard)

Metric	Value
Recall	1.0 (all 10 degrading engines detected)
Precision	0.23
F1	0.37
Avg lead time	195 cycles before failure
Detection rate	10/10 engines

High recall / lower precision is a deliberate design choice for a safety-critical path — false negatives (missing a failing engine) are more costly than false positives (unnecessary inspection). See FAILURE_MODES.md for the full trade-off rationale.

Encryption overhead

Operation	Mean	P95
Encrypt	0.019 ms	0.027 ms
Decrypt	0.018 ms	0.024 ms

Full JSON: reports/benchmark_results.json

---

Known limits

• Threat model covers data ingestion and API surface; training infrastructure compromise is not in scope
• RBAC is application-layer only — no hardware isolation, TEE, or confidential compute
• Audit log is append-only with hash-chain integrity; not blockchain-sealed
• Anomaly gate is tuned for FD001; other C-MAPSS subsets (FD002–FD004) need recalibration
• Network resilience drops below target under 10%+ packet loss — documented in benchmark (target_met: false)
• LSTM/Transformer ensemble not yet evaluated against adversarial sensor injection beyond Gaussian noise

---

Setup

git clone https://github.com/poojakira/PulseNet-RUL-Forecasting
cd PulseNet-RUL-Forecasting
pip install -r requirements.txt

# Run pipeline
python main_pipeline.py

# Start API
uvicorn src.pulsenet.api.app:app --reload

# POST /predict with sensor payload
# Audit log written to logs/audit.jsonl

# Run security tests
pytest tests/test_security.py -v

# Reproduce benchmarks
python scripts/run_validation.py

Docker:

docker-compose up

---

References

• NASA C-MAPSS Jet Engine Simulated Dataset
• NIST AI Risk Management Framework (AI RMF 1.0)
• NIST SP 800-204D: DevSecOps for Microservices
• STRIDE Threat Modelling — Microsoft
• MITRE ATLAS — Adversarial Threat Landscape for AI Systems