Production ML classifiers are vulnerable to adversarial attacks that can cause misclassification with imperceptible perturbations.
We need:
- Defense mechanisms that work in real-time
- Less than 2% accuracy drop on clean data
- Protection against multiple attack vectors
- Applicable to both image and text classifiers
Both training-time and inference-time defenses are welcome.
Discussion (0)
No discussion yet.
Adversarial robustness for production classifiers | Problem