Explain the privacy issues raised by artificial intelligence systems : Poisoning Attacks, Privacy, and Backdoor Attacks

The video tutorial covers three main types of attacks on machine learning models: poisoning, privacy, and backdoor attacks. Poisoning attacks involve injecting malicious data into the training dataset to alter the model's decision boundary. Privacy attacks focus on breaking confidentiality and include inference attacks that can occur during both training and production stages. Backdoor attacks aim to introduce hidden behaviors in models that persist even after retraining. The tutorial highlights the complexity and current research gaps in these areas, emphasizing the need for further exploration and development of protection measures.

What is the primary goal of a poisoning attack on a machine learning model?

• To change the input example

• To maintain behavior after retraining

• To extract data from the model

• To modify the decision boundary

Which attack strategy involves modifying only the labels in a supervised learning dataset?

• Data injection

• Data modification

• Logic corruption

• Label modification

What is the main objective of privacy attacks?

• To break confidentiality

• To modify decision boundaries

• To maintain backdoors

• To inject malicious data

Which type of attack helps to extract specific data from a model?

• Logic corruption attack

• Model inversion attack

• Data modification attack

• Label modification attack

What is a key characteristic of backdoor attacks?

• They operate even after retraining

• They change the input example

• They require full access to training data

• They modify only the labels

Why is it challenging for small companies to detect backdoors?

• They can modify decision boundaries

• They can inject any data

• They have full access to training data

• They lack computing power

What is a current challenge in the research of poisoning, backdoors, and privacy attacks?

• They are easy to detect

• They are deeply analyzed

• They are not deeply analyzed

• Good protection measures exist