Use a real-life example of an AI system to discuss some impacts of cyber attacks : Poisoning Attacks, Privacy, and Backd

The video tutorial explores various types of attacks on machine learning models, including poisoning, privacy, and backdoor attacks. Poisoning attacks involve injecting malicious data to alter model behavior, while privacy attacks focus on extracting confidential information. Backdoor attacks aim to embed hidden functionalities that persist even after retraining. The tutorial highlights the challenges in detecting these attacks and the need for further research in this area.

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

• To improve data privacy

• To speed up the training process

• To change the classification boundary

• To enhance the model's accuracy

Which strategy allows an adversary to modify only the labels in a supervised learning dataset?

• Data injection

• Logic corruption

• Label modification

• Data modification

What is a model extraction attack?

• An attack to corrupt model logic

• An attack to extract data from a model

• An attack to improve model efficiency

• An attack to modify training data

How can privacy attacks compromise confidentiality?

• By enhancing data encryption

• By retrieving information through inference attacks

• By improving model accuracy

• By reducing data redundancy

What is the main objective of a backdoor attack?

• To inject additional behavior that persists after retraining

• To enhance data security

• To improve the model's performance

• To remove existing vulnerabilities

Why are backdoors difficult to detect in neural networks?

• They are only present in outdated models

• They are hidden within a small set of neurons

• They require large datasets to be visible

• They are easily removed during retraining

What makes it challenging for small companies to detect backdoors?

• Lack of computing power to retrain models

• Inability to access public models

• Limited data storage capacity

• Insufficient training data