@ericyocam6Shipped: Adversarial AI — Compromised Models and Attack Datasets
An open-source collection of adversarial AI research artifacts developed for Book 2 of The Applied AI Universe Coding Guide series.
The repository contains deliberately compromised machine learning models and datasets used to demonstrate how adversarial attacks can undermine AI systems.
Research artifacts include poisoned training datasets, corrupted model parameters, manipulated symbolic AI components, and backdoored LoRA adapters.
These examples support the study of adversarial machine learning, data poisoning, model integrity, backdoor vulnerabilities, and AI security evaluation.
The collection is designed to work alongside the clean baseline models from Book 1 and the corresponding defensive experiments in Book 3, allowing researchers and educators to examine the complete progression from model development to attack and defense.
IMPORTANT: These artifacts are intentionally compromised and must not be deployed in production or loaded into trusted systems. They are intended for authorized research and controlled educational environments.
Comments2
Are corporate security teams using these poisoned datasets to train staff, or is the audience mostly professors and students working through the books?
Great question, Ryan! The primary audience right now is educators, students, and AI security researchers using the artifacts to understand how adversarial attacks affect machine learning systems.
That said, I also see potential value for corporate security teams, particularly in training AI developers, red teams, and security analysts to recognize data poisoning, model backdoors, and integrity risks.
I haven't established widespread corporate adoption, but the artifacts could support hands-on security training in controlled environments. The goal is to make adversarial AI concepts practical and accessible beyond the classroom. Appreciate the question!
2 comments
Are corporate security teams using these poisoned datasets to train staff, or is the audience mostly professors and students working through the books?
Great question, Ryan! The primary audience right now is educators, students, and AI security researchers using the artifacts to understand how adversarial attacks affect machine learning systems.
That said, I also see potential value for corporate security teams, particularly in training AI developers, red teams, and security analysts to recognize data poisoning, model backdoors, and integrity risks.
I haven't established widespread corporate adoption, but the artifacts could support hands-on security training in controlled environments. The goal is to make adversarial AI concepts practical and accessible beyond the classroom. Appreciate the question!