Evaluate the impact of privacy issues, cyberattacks, and malware on your AI application : Comparing Defence Approaches a

The video tutorial discusses various metrics used to evaluate model defenses, focusing on protected model accuracy, classification accuracy variance (CAV), and QQ and Co values. It highlights the importance of these metrics in assessing the effectiveness of defenses against adversarial inputs. The tutorial also explains how to run a security evaluation script to test defenses against different attacks, emphasizing the significance of thermometer encoding as a superior defense. The session concludes with a homework assignment to choose the best defense based on collected metrics.

Which defense method was found to improve model accuracy?

• Gradient Masking

• Random Noise Injection

• Thermometer Encoding

• Adversarial Training

What does a positive classification accuracy variance (CAV) indicate?

• The model performs worse on adversarial examples

• The model has improved accuracy on non-adversarial examples

• The model's accuracy is unchanged

• The model's accuracy is worse on non-adversarial examples

What is the role of classification rectify ratio in CAV?

• It measures the accuracy on adversarial examples

• It indicates the percentage of examples misclassified by the secure model

• It represents the percentage of examples correctly classified by the secure model

• It shows the similarity of model responses

What is the purpose of the security evaluation script?

• To test defenses against adversarial attacks

• To improve model accuracy

• To generate adversarial examples

• To calculate classification accuracy variance

Which defense showed zero accuracy against a specific attack?

• Adversarial Training

• RC

• EIT

• Thermometer Encoding