Failures

Adversarial Input

1 min read

What It Means

Carefully crafted inputs designed to cause AI systems to make mistakes they wouldn't make on normal inputs.

Adversarial inputs exploit model vulnerabilities, often appearing innocuous to humans while causing system failures.

Examples

  • Typos that fool classifiers
  • Semantic-preserving perturbations
  • Out-of-distribution triggers
  • Multi-modal attacks

Robustness

Systems should be tested against adversarial inputs, not just normal cases.

failuressecurityrobustness