@ericyocam6Professor, California Polytechnic State University
Shipped: Applied AI Universe Coding Guide
A three-book, code-first series that goes from building neural networks to generative AI, then attacking those models, then defending them with adversarial training and certified robustness.
machine-learningdeep-learninggenerative-aiadversarial-mlai-securityquantum-mleducation
Comments2
The jump from attacking to defending is where most tutorials stop short. With adversarial training versus certified robustness, do your readers see the accuracy and compute tradeoff in the code, say PGD training against randomized smoothing certificates on the same model?
Great question, Ryan. That's exactly the kind of tradeoff the series is designed to explore. Readers build models in Book 1, attack those same models in Book 2, and implement defenses in Book 3 using executable Google Colab notebooks. We cover adversarial training and certified robustness, with measured results rather than theoretical claims. Importantly, we also retest defenses against adaptive attackers to distinguish apparent robustness from actual protection. The goal is for readers to understand not just whether a defense works, but what it costs in accuracy, computation, and overall model performance. A direct PGD-versus-randomized-smoothing comparison on identical model configurations would be a useful additional benchmark.
2 comments
The jump from attacking to defending is where most tutorials stop short. With adversarial training versus certified robustness, do your readers see the accuracy and compute tradeoff in the code, say PGD training against randomized smoothing certificates on the same model?
Great question, Ryan. That's exactly the kind of tradeoff the series is designed to explore. Readers build models in Book 1, attack those same models in Book 2, and implement defenses in Book 3 using executable Google Colab notebooks. We cover adversarial training and certified robustness, with measured results rather than theoretical claims. Importantly, we also retest defenses against adaptive attackers to distinguish apparent robustness from actual protection. The goal is for readers to understand not just whether a defense works, but what it costs in accuracy, computation, and overall model performance. A direct PGD-versus-randomized-smoothing comparison on identical model configurations would be a useful additional benchmark.