adversarial examples

Adversarial examples are inputs designed to fool models into making incorrect predictions. **For vision**: Imperceptible pixel perturbations cause misclassification (panda → gibbon). **For NLP**: Character swaps ("g00d"), word substitutions, paraphrase attacks, prompt injections. **Attack types**: **White-box**: Attacker has model access, uses gradients (FGSM, PGD). **Black-box**: Query-only access, transfer attacks, search-based. **Targeted vs untargeted**: Force specific wrong output vs any error. **NLP challenges**: Discrete tokens (can't use gradients directly), semantic constraints (must remain meaningful). **Techniques**: TextFooler, BERT-Attack, word substitution, character-level perturbations. **Why they exist**: Models rely on spurious features, decision boundaries are brittle, high-dimensional input spaces. **Real-world impact**: Spam evasion, content moderation bypass, autonomous vehicle attacks, biometric spoofing. **Defenses**: Adversarial training, input preprocessing, certified robustness, ensemble methods. **Detection**: Identify adversarial inputs before classification. Critical security concern for deployed ML systems.