What is adversarial training (PGD, FGSM)?
hardAnswer
- Generate small adversarial perturbations that fool the current model (FGSM = one gradient step; PGD = multiple projected gradient steps), then include them in training with the true labels.
- Improves robustness to bounded input perturbations at ~2-10x compute overhead.
- Trade-off: clean accuracy typically drops 3-5% while adversarial accuracy jumps from ~0% to 40-60%.
- Standard baseline for robust ML.
Check yourself — multiple choice
- Adversarial training removes gradients
- Train on adversarial perturbations (FGSM/PGD) with true labels — robustness at compute cost, some clean-accuracy loss
- Same as data augmentation with rotations
- Adversarial training improves clean accuracy
PGD/FGSM adversarial training: worst-case perturbations in the training set → robustness.
#regularization#training
Practise Deep Learning
214 interview questions in this topic.
Related questions
- How does dropout work and when is it applied?
- What is stochastic depth / DropPath and where is it used?
- What is early stopping and how do you configure it?
- What is Stochastic Weight Averaging (SWA)?
- What is an EMA of weights and why do modern training recipes use it?
- Your validation loss is lower than your training loss. Is something broken?