Your threat models name the attackers. This module shows what they actually do to a machine learning model. Adversarial machine learning is the study of deliberately making a model fail, leak, or be copied, and it splits into a few attack families worth knowing by name.
Evasion attacks craft inputs at prediction time so the model misclassifies them. Data poisoning corrupts the training data so the model learns the wrong behavior, sometimes with a hidden backdoor. Model extraction copies a model by querying it, and privacy attacks such as membership inference and training data extraction pull out information about what the model was trained on.
Watch the lessons in order, then continue on to LLM & Agent Attacks, where these ideas meet language models and autonomous tools.
An overview of adversarial machine learning: how attackers poison data sets before training and evade models after deployment.
How an attacker who modifies training data can control a model's behavior at test time, including a demonstration against a commercial AutoML platform.
A research talk on membership inference: how an attacker can tell whether a particular data point was used to train a model, and why that is a privacy risk.
A walkthrough of the research showing that querying a language model can recover individual examples it memorized during training.