RAIL: Risk-Aware Human-in-the-Loop Framework with Adaptive Intrusion Response...
summary
RAIL is a runtime framework that fuses heterogeneous cues into an Intrusion Risk Index (IRI) using a weighted Noisy‑OR over curvature actuation integrity, time‑to‑collision proximity, and observation‑shift consistency, then uses a contextual bandit to pick cue‑specific shields and a graded authority blend (alpha threshold ≈ 0.3) between nominal and safeguarded controls. Implemented in MetaDrive and CARLA with risk‑prioritized replay and dual reward shaping, it matches expert success while reduces attack success rates by up to 48% and lowers operator disengagement under attack by >50%.
AI-written summary · about
chapters (9)
- 0:00RAIL framework intro
- 1:36Limitations of safety approaches
- 3:25Risk as first-class variable
- 4:48System model and risk metrics
- 6:15RAIL architecture and IRI calculation
- 9:25Graded authority and safety learning
- 11:10Evaluation setup and simulation
- 12:55Results under attacks
- 14:43Conclusion and future work
related talks
30mRisk-Driven Fault Injection for Fast & Furious
38mSecurity for AI Agents Using an Ensemble of Fine-tuned... (Lidan Hazout, Bar Kaduri)
31mAI Interpretability as a Security Control: Introducing the CIRCUIT Framework
43mRDI: A Free and Easy Framework for Ransomware Readiness Assessment
20mAI Red Teaming and Automation: Exploring Societal Risks in...
28mAI Under the Hood: Unmasking Hidden Threats
17mGuardrails First: Designing AI Agents for Security Operations
34mRapid Risk Assessment: A Lightweight Approach