Evaluate the impact of privacy issues, cyberattacks, and malware on your AI application : Testing Practical Defence from

The video tutorial discusses the concept of adversarial inputs and how they can be processed before being fed into a model. It introduces a defense mechanism called input transformation, which involves modifying images through techniques like cropping and compression. The tutorial evaluates the effectiveness of this defense, noting that while it is good against attacks, it may not be suitable for normal inputs. The importance of maintaining classification accuracy is emphasized, and the video concludes with a discussion on different defenses and metrics.

What is the primary goal of modifying an image with adversarial pixels?

• To reduce the file size of the image

• To test the robustness of image processing models

• To enhance the image quality

• To change the image's appearance to the human eye

Which of the following is NOT a technique proposed in the white paper for modifying images?

• Rescaling

• Color inversion

• Compression

• Cropping

What is a significant drawback of the defense mechanism discussed in the video?

• It is too simple to implement

• It is ineffective against adversarial attacks

• It reduces the accuracy on normal inputs

• It requires expensive hardware

Why is maintaining classification accuracy on normal testing examples important?

• To ensure the model is fast

• To ensure the model's performance is reliable

• To avoid overfitting

• To keep the model's predictions consistent

What does the term 'classification accuracy variance' refer to?

• The difference in accuracy between different models

• The fluctuation in accuracy due to random noise

• The change in accuracy over time

• The variation in accuracy between adversarial and normal examples