A Study on Extending the TARA Methodology for Physical AI: A Case Study of ADS
summary
Extends STRIDE by merging MITRE ATLAS and OWASP Top 10 for LLMs into a unified TARA that normalizes 170 sources into 89 attack scenarios and generates 595 threat scenarios applied to an ADS reference stack on NVIDIA Jetson. Produces 38 security checklists mapped to all 32 UN R155 Annex 5 Part A categories, claims 29% more AI-specific risks than STRIDE alone, achieves ML 3 for data‑source integration, and derives verifiable checks from an Attack Library built from 42 CWEs, 76 CVEs, and 47 papers.
AI-written summary · about
chapters (19)
- 0:00Extending TARA for Physical AI
- 1:16Conventional threat modeling limits
- 2:22Research approach directions
- 3:30Expert feedback and design input
- 3:53Attack library and threat mapping
- 5:41ADS architecture case study
- 6:48170 attack sources to 89 scenarios
- 7:31595 threat scenarios identified
- 8:12STRIDE threats on software assets
- 8:32AI threats via MITRE ATLAS
- 8:54LLM threats via OWASP AI Top 10
- 9:15Damage scenarios by ISO 21434
- 9:38Attack paths to vehicle consequences
- 10:19Security requirements from threats
- 10:4338 security checklist items
- 11:28Validation vs UN Reg 155 Annex 5
- 12:11STRIDE vs extended analysis
- 13:16Data integration maturity level
- 13:59Summary and future work
related talks
28mAI Under the Hood: Unmasking Hidden Threats
26mImproving Cybersecurity Posture with an AI Security Assistant
33mSecuring Agentic AI With PyTorch: Threat Modeling & LLM Red Teaming in Practice
30mNavigating the New AI Security Landscape
25mThreat Modeling for AI Apps with Attacks as Code
50mThe ART of Probable: Test with AI, Atomic Red Team, and Threat Metrics
39mThreat Modeling Using LLMs
13mAPTGen: An Approach towards Generating Practical Dataset Labelled with Targeted...