GDG on Campus Texas Tech University - Lubbock, United States
🚀 Launching Attacks – Running a Greedy Coordinate Gradient (GCG) attack on Llama 3.2 3B and Phi 3.5 Mini Instruct to extract unsafe responses. 🔍 Exposing Vulnerabilities – Identifying flaws in response generation and safety mechanisms of these models. 🛠️ Reinforcing Defenses – Fine-tuning both models using adversarial examples to resist future attacks. 💡 Testing the Fixes – Reassess security
1 RSVP'd
Fine-Art of Fine-Tuning: Attacking and Securing LLMs
● Perform a Greedy Coordinate Gradient
(GCG) attack on Llama 3.2 3B and Phi 3.5
Mini Instruct to extract malicious
responses
● Identify weaknesses in the response
generation and safety mechanisms of
both models
● Train Llama 3.2 3B and Phi 3.5 Mini Instruct
using adversarial examples to resist
prompt attacks
● Reassess model security by attempting
new prompt attacks post-fine-tuning
PETR 110; 5:30 PM to 6:30 PM, Feb 13 & Feb 20
February 13 – 14, 2025
11:30 PM – 12:30 AM (UTC)
Organizer
Contact Us
