WIP: From Detection to Explanation: Using LLMs for Adversarial Scenario Analysis...
summary
Claims a framework that turns structured DSL vehicle scenarios into interpretable anomaly explanations using LLMs, evaluates 40 MUTCD-based cases, and requires no labeled training data. Uses AutoSec-X: 40 MUTCD scenarios, passes DSL-encoded scenes through Llama/Gemini/Shivven with zero-shot CoT prompting, scores outputs with ROUGE/BLEU/ESBERT plus human review, and finds Gemini models outperform smaller models with higher semantic alignment and frequent citation of MUTCD sections.
AI-written summary · about
related talks
47mHarnessing LLMs for Next-Level Security Analysis
28mExplaining Machine Learning Models with LLMs
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
13mIntegrating Large Language Models into Security Incident Response
39mThreat Modeling Using LLMs
33mExplore LLMs and Multimodal Approaches in Test Automation
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
01h 49mAI engineering with open access LLMs that lie, curse, and steal