aselimc Threat Modeling
STRIDE Methodology
| Threat | Description | Mitigation |
|---|---|---|
| Spoofing | Pretending to be another user/service | AuthN, certificates |
| Tampering | Modifying data in transit/at rest | Integrity checks, signing |
| Repudiation | Denying actions | Audit logs |
| Information Disclosure | Unauthorized data access | Encryption, access control |
| Denial of Service | Making service unavailable | Rate limiting, scaling |
| Elevation of Privilege | Gaining unauthorized access | Least privilege, RBAC |
Process
- Identify attack surfaces and trust boundaries
- Map data flows between components
- Apply STRIDE to each data flow
- Rate severity: likelihood x impact (Critical/High/Medium/Low)
- Prioritize mitigations by severity
ML-Specific Threats
- Model extraction: API queries to reconstruct model
- Adversarial inputs: crafted inputs to fool model
- Data poisoning: corrupted training data
- Prompt injection: malicious input to LLMs
Tools
OWASP Threat Dragon, Microsoft Threat Modeling Tool