Stanford CS230 | Autumn 2025 | Lecture 4: Adversarial Robustness and Generative Models | Stanford Online

In this lecture, Kian Katanforoosh explores two main topics: adversarial robustness and generative modeling. The discussion on adversarial robustness covers attacks on AI models, including prompt injection and data poisoning, and the importance of building proactive defenses. Katanforoosh outlines three waves of adversarial attacks, highlighting how models are increasingly vulnerable due to their reliance on instructions and context. The lecture then transitions to generative models, focusing on GANs and diffusion models, which are used in image and video generation. Katanforoosh explains the differences between discriminative and generative models, emphasizing the latter's ability to learn the underlying distribution of data. The session includes interactive Q&A, addressing concerns about the sensitivity of neural networks to forged images and potential defenses against attacks.

Outlines

Part 1: Introduction to Attacks

Part 2: Generative Adversarial Networks (GANs)

Part 3: Diffusion Models

Sign in to continue reading, translating and more.

Continue

Stanford CS230 | Autumn 2025 | Lecture 4: Adversarial Robustness and Generative Models

Stanford Online

Part 1: Introduction to Attacks

Introduction to Adversarial Robustness and Generative Modeling

Adversarial Attacks: Forging Images and Understanding Model Vulnerability

Advanced Adversarial Attacks and Patch Optimization

Neural Network Sensitivity and Fast Gradient Sign Method

Defenses Against Adversarial Attacks and Backdoor Attacks

Prompt Injection Attacks and Introduction to Generative Modeling

Part 2: Generative Adversarial Networks (GANs)

Generative Models: GANs and the Minimax Game

GAN Training Losses and the Minimax Game

GAN Training Challenges and Linearity in Code Space

Part 3: Diffusion Models

Introduction to Diffusion Models and the Forward Diffusion Process

Denoising and the Training Process for Diffusion Models

Sampling and Test Time Inference with Diffusion Models

Latent Diffusion and Video Generation

Stanford CS230 | Autumn 2025 | Lecture 4: Adversarial Robustness and Generative Models

Stanford Online

Part 1: Introduction to Attacks

00:05Introduction to Adversarial Robustness and Generative Modeling

Introduction to Adversarial Robustness and Generative Modeling

07:14Adversarial Attacks: Forging Images and Understanding Model Vulnerability

Adversarial Attacks: Forging Images and Understanding Model Vulnerability

16:06Advanced Adversarial Attacks and Patch Optimization

Advanced Adversarial Attacks and Patch Optimization

24:21Neural Network Sensitivity and Fast Gradient Sign Method

Neural Network Sensitivity and Fast Gradient Sign Method

32:40Defenses Against Adversarial Attacks and Backdoor Attacks

Defenses Against Adversarial Attacks and Backdoor Attacks

41:52Prompt Injection Attacks and Introduction to Generative Modeling

Prompt Injection Attacks and Introduction to Generative Modeling

Part 2: Generative Adversarial Networks (GANs)

49:40Generative Models: GANs and the Minimax Game

Generative Models: GANs and the Minimax Game

57:31GAN Training Losses and the Minimax Game

GAN Training Losses and the Minimax Game

1:06:04GAN Training Challenges and Linearity in Code Space

GAN Training Challenges and Linearity in Code Space

Part 3: Diffusion Models

1:15:35Introduction to Diffusion Models and the Forward Diffusion Process

Introduction to Diffusion Models and the Forward Diffusion Process

1:25:54Denoising and the Training Process for Diffusion Models

Denoising and the Training Process for Diffusion Models

1:34:01Sampling and Test Time Inference with Diffusion Models

Sampling and Test Time Inference with Diffusion Models

1:42:28Latent Diffusion and Video Generation

Latent Diffusion and Video Generation