WeaknessASC-2026-0064Other weakness
Adversarial examples against ML classifiers
LowWorkarounds only
Small, crafted changes to inputs make spam, fraud, malware or content classifiers give the wrong answer while the input still looks normal.
How to detect it
- Classifier decisions that flip on near-identical inputs
- Attack traffic tuned against your classifier
Workarounds
- Combine ML decisions with rules and other signals
- Train with adversarial examples and monitor drift
Robustness to adversarial examples remains an open research problem.
How to investigate
- Collect the evading samples and retrain
- Add rule-based checks for the pattern