security_architecture

---

iso_doc_type: "Description" iso_viewpoint: "SecurityView" type: "security" title: "Security Architecture & Threat Mitigation" description: "ISO 42010 SecurityView / ISO 15289 Description documentation for STRIDE threat modeling, model signing, and rate limiting." tags: ["iso42010", "okf", "security_view", "security", "threat_model"] timestamp: "2026-10-01T17:50:00Z" generated: "agent:wiki-iso-documentation-standard" verified: "true" last_verified_commit: "073b5fd"

Security Architecture & Threat Mitigation

Purpose: Detail the threat analysis, defense-in-depth security controls, input validation gates, and cryptographic protections implemented across the system.

---

1. STRIDE Threat Model & Defense Matrix

Threat Category	Potential Attack Vector	System Defense Mechanism	Source File Citation
Spoofing	Unauthorized clients submitting fake predictions	JWT / Bearer token validation, mTLS in container overlay	controller/kafka_app.py:L69-L79
Tampering	Model artifact poisoning or payload modification	SHA-256 hash attestation via InferSigner	utils/signers.py:L21-L51
Repudiation	Disputing historical inferences or training runs	Immutable MLflow run history and signed prediction logs	io/services.py:L162-L252
Information Disclosure	Leakage of PII or dataset values into application logs	Loguru message sanitization, log-leakage regression tests	tests/controller/test_log_leakage.py:L1-L60
Denial of Service (DoS)	Exhaustion of compute via unbounded payloads or floods	Sliding-window RateLimiter & max column checks (MAX_TRACKED_IPS)	controller/kafka_app.py:L82-L112
Elevation of Privilege	Code injection via untrusted configuration serialization	OmegaConf strict YAML schema parsing (no eval / pickle)	io/configs.py:L16-L68

---

2. Sliding-Window Rate Limiting (controller/kafka_app.py)

To protect inference endpoints against volumetric Denial of Service (DoS), the RateLimiter enforces request quotas per client IP:

graph TD
    REQ["Incoming Request (Client IP)"] --> CHECK{"Request Count in Window < 100?"}
    CHECK -->|Yes| ALLOW["Record Timestamp & Allow Request"]
    CHECK -->|No| REJECT["Raise 429 Too Many Requests"]

    subgraph MemoryCleanup["Memory Bound Protection"]
        EVICT["Evict oldest IPs when tracked > 10,000"]
    end
    ALLOW -.-> EVICT

Loading

• Algorithm: In-memory sliding window using double-ended queues (collections.deque).
• Quota: Default 100 requests per 60-second window.
• Resource Bound: Upper bound on tracked IP cache (MAX_TRACKED_IPS = 10,000) prevents memory exhaustion attacks.

---

3. Cryptographic Model Attestation (utils/signers.py)

The InferSigner computes a deterministic SHA-256 digest over serialized input-output pairs:

$$ ext{Signature} = ext{SHA256}( ext{JSON}( ext{Inputs}) \mathbin{\Vert} ext{JSON}( ext{Outputs}))$$

This guarantees that inference results can be cryptographically verified against historical records during external compliance audits.

---

Related: HITL Governance · Verification Triad · Master Index