Mis on vasturikkursus ja milline on selle kaubanduslik kompromiss?

Adversarial training is a method to increase the robustness of machine learning models against attacks where input data is intentionally modified to deceive the model (adversarial examples).

The essence of adversarial training is that during training, specially generated adversarial examples are added to normal data, and the model learns to classify them correctly. This increases its resistance to similar attacks in the future.

Trade-off of adversarial training:

• Advantages:

  • Significantly increases the security and reliability of the model.
  • Improves resistance to small data distortions.

• Disadvantages:

  • Increases training time and resources.
  • May reduce accuracy on clean (unmodified) data.
  • Generating high-quality adversarial examples can be complex.

Thus, adversarial training is a balance between enhancing the model's security and the training costs, as well as potential overall performance degradation.