Machine Learning Security

Scoreboard · Submit Flags

Friday Survey

Scores archived 5-5-24

Presentation: AI Workshop Summary

KEY · PPTX· PDF

Attack References

It’s disturbingly easy to trick AI into doing something deadly
GhostStripe attack haunts self-driving cars by making them ignore road signs
MadRadar hack can make self-driving cars 'hallucinate' imaginary vehicles and veer dangerously off course
Two big computer vision papers boost prospect of safer self-driving vehicles

Defenses

ML 140: Deep Neural Rejection (45 pts extra)

Large Language Models

ML 120: Bloom LLM (15 pts + 15 extra)
ML 121: Prompt Engineering Concepts (20 pts)
ML 122: Comparing LLMs on Colab (10 pts + 10 extra)
ML 123: Running Llama 3 Locally (15 pts extra)
ML 124: Evaluating an LLM with Trulens (15 pts extra)
ML 126: Building RAGs (15 pts extra)
ML 127: Encoding Text with BERT (10 pts extra)
ML 128: Using AnythingLLM to Embed Custom Data (10 pts extra)
ML 129: Embedding Words with BERT (40 pts)

ML 125: Jupyter Notebook on a Mac M1 (10 pts extra)

Generating Code

ML 160: GitHub Copilot (15 pts extra)
ML 131: Generating Python Code with Gemini (40 pts extra)
Violent Python Challenges (extra)

References

SecML: Secure and Explainable Machine Learning in Python
ChatGPT Prompt Engineering for Developers
Prompt Engineering Guide
Google's Generative AI learning path
A jargon-free explanation of how AI large language models work
Pinecone Makes Accurate, Fast, Scalable Generative AI Accessible to Organizations Large and Small with Launch of its Serverless Vector Database
Pinecone Vector Database
Free Training Building Applications with Vector Databases

The Databricks Data Intelligence Platform
Attention in transformers, visually explained

Machine Learning Security

Scoreboard · Submit Flags Friday Survey Scores archived 5-5-24

Presentation: AI Workshop Summary

Understanding Prompts

Google Learning

Security Risks

Awareness: Demonstrating Capabilities

Technical: Inner Components

Attacks