@ericyocam6Shipped: The Applied AI Universe — Defending 35 AI Systems
Book 3 of The Applied AI Universe Coding Guide is an open-source, hands-on educational resource examining how to defend AI models against adversarial attacks, and how those defenses perform when attackers adapt.
The framework includes 35 classical and hybrid/quantum defense experiments covering machine learning, deep neural networks, generative AI, reinforcement learning, natural language processing, and quantum computing.
Each experiment reconstructs an adversarial attack from Book 2, applies a corresponding defensive technique, measures performance, and then evaluates the defense against an adaptive attacker who knows the protection is present.
Defense techniques include adversarial training, input validation, anomaly detection, differential privacy, cryptographic integrity verification, data provenance, and parameter recovery.
Across 29 defenses with comparable measurements, the reported average improvement decreased from 0.271 under the original attacks to 0.191 under adaptive attacks. The framework intentionally includes ineffective and partially successful defenses to help learners understand why security controls sometimes fail.
Built as reproducible Google Colab notebooks, the project provides researchers, educators, and AI developers with practical examples of defensive machine learning, robustness evaluation, and secure AI engineering.
These are educational research experiments, not certified production defenses or evidence of a general quantum advantage.
Comments2
Testing each defense a second time against an attacker who knows it's there, and showing the average gain drop from 0.271 to 0.191, is the honest part most security courses skip. Students who work through the failed and half-working defenses will start their first job knowing that a fix that holds up in a classroom can still slip once a real attacker adjusts.
Thanks, Ryan! That's exactly the lesson I wanted students and researchers to take away from Book 3. A defense that performs well against a known attack isn't necessarily robust when an adversary adapts. The drop from 0.271 to 0.191 illustrates why evaluating defenses against adaptive attacks is so important. I deliberately included defenses that failed or only partially succeeded because those outcomes can be just as educational as successful ones. My goal with the three-book series is to take learners through building AI models, attacking them, and then defending them under increasingly realistic conditions. Ultimately, understanding why a defense fails is an important part of learning how to build more secure AI systems. Appreciate your perspective!
2 comments
Testing each defense a second time against an attacker who knows it's there, and showing the average gain drop from 0.271 to 0.191, is the honest part most security courses skip. Students who work through the failed and half-working defenses will start their first job knowing that a fix that holds up in a classroom can still slip once a real attacker adjusts.
Thanks, Ryan! That's exactly the lesson I wanted students and researchers to take away from Book 3. A defense that performs well against a known attack isn't necessarily robust when an adversary adapts. The drop from 0.271 to 0.191 illustrates why evaluating defenses against adaptive attacks is so important. I deliberately included defenses that failed or only partially succeeded because those outcomes can be just as educational as successful ones. My goal with the three-book series is to take learners through building AI models, attacking them, and then defending them under increasingly realistic conditions. Ultimately, understanding why a defense fails is an important part of learning how to build more secure AI systems. Appreciate your perspective!