Explain the negative impacts of artificial intelligence systems on society : Adversarial Attacks Metrics and White-Box A

The video tutorial covers key metrics for evaluating adversarial attacks, focusing on misclassification, imperceptibility, robustness, and speed. It explains how these metrics are used to compare different attacks, highlighting the importance of misclassification rate, average confidence, and imperceptibility measures like LP distortion and structural similarity. Robustness is discussed in terms of noise tolerance and image-specific measures, while speed is noted for its relevance in recent attacks. The tutorial concludes with a summary and a preview of future content on practical execution of attacks.

Which metric is considered the most common criterion for comparing adversarial attacks?

• Imperceptibility

• Misclassification

• Robustness

• Speed

What does the misclassification ratio measure in adversarial attacks?

• The percentage of examples misclassified

• The speed of computation

• The robustness to noise

• The imperceptibility of modifications

Which measure is more consistent with human visual perception when evaluating imperceptibility?

• Average LP distortion

• Average structural similarity

• Noise tolerance estimation

• Robustness to Gaussian blur

What does noise tolerance estimation reflect in the context of adversarial attacks?

• The structural similarity of the attack

• The amount of noise an example can tolerate

• The speed of the attack

• The imperceptibility of the attack

Why is computational speed important in recent adversarial attacks?

• It determines the imperceptibility of the attack

• It affects the robustness to image compression

• It allows attacks to benefit from GPU support

• It increases the misclassification rate