ManySpace
Türkçe Sign in Sign up

AI & MLRecommendedAIML-SEC-003

Data Poisoning and Backdoors

Corrupting training data so a model fails generally, or behaves normally except on a hidden trigger.

Difficulty
3 / 5
Importance
4 / 5
Study time
1–2 h
Kind of work
Mixed

Why it matters

The learner can vary the poisoned fraction of a training set and predict how a backdoor's success and clean accuracy change.

The route to this topic

Route AIML-SEC-003
What Is a Model? Learning from Data Features and Targets Datasets and Splits Linear Regression I: Prediction and Residuals Sampling and Splits Train, Validation, and Test Errors Loss and Fitting: Squared Error Distribution Shift Offline Evaluation Objective Functions and Empirical Risk Robustness Evaluation Training vs Serving Adversarial Examples Threat Modeling for ML Systems Data Poisoning and Backdoors