neural-network-adversarial-vulnerability-general

IN derived (depth 1)

Created 2026-06-21T09:59:01+00:00 · Reviewed 2026-06-21T15:37:01+00:00

Adversarial vulnerability is a general property of neural networks spanning supervised learning (imperceptible image perturbations), reinforcement learning (shared adversarial features across MDPs), and even single-pixel attacks — not a quirk of any particular architecture or domain.

Justifications

SL — Adversarial fragility manifests across classification, RL, and minimal perturbations

Antecedents (all must be IN):

  • IN dl-adversarial-examples-imperceptible-perturbations — Adversarial examples are inputs with small, human-imperceptible perturbations that cause neural networks to confidently misclassify them.
  • IN deep-rl-adversarial-vulnerability — Deep RL policies share adversarial features across MDPs (Korkmaz 2022), making adversarial vulnerability a key limitation of deep RL
  • IN ml-adversarial-single-pixel — Adversarial examples are deliberately crafted input perturbations that can cause misclassification, sometimes by changing as little as a single pixel

Dependents

These beliefs depend on this one: