This Tiny Change BREAKS AI 🤯 | FGSM Adversarial Attack Explained скачать в хорошем качестве

This Tiny Change BREAKS AI 🤯 | FGSM Adversarial Attack Explained

NOTEBOOK: https://colab.research.google.com/dri... LATEX FGSM DOCUMENT: https://www.overleaf.com/read/qpjjnnb... 00:00:00 – Introduction to FGSM Attack Overview of fast gradient sign method and adversarial goal (6→8 example) 00:00:45 – Loss Function & Gradient Descent Review Explanation of J(θ, X, Y), θ parameters, and training vs. attack 00:01:30 – FGSM Formula Breakdown Derivation of X′ = X + ε·sign(∇ₓJ(θ, X, Y)) and why it works 00:03:37 – Tuning Epsilon (ε) Values Impact of small vs. large ε on perturbation visibility and model fooling 00:04:30 – Setting Up Google Colab & Imports PyTorch, transforms, and MNIST dataset libraries 00:06:00 – Defining a Simple CNN Architecture Building a two-layer conv + one fully connected network in PyTorch 00:08:10 – Loading MNIST & Preparing for FGSM Transforming images to tensors and initializing the model for evaluation 00:10:00 – Implementing the FGSM Attack Loop Computing gradients, creating perturbed images, and checking misclassification 00:14:48 – Visualizing Adversarial Examples & Results Displaying original vs. perturbed images and classification changes 00:15:17 – Conclusion & Further Questions Summary of FGSM’s impact and invitation for audience questions