liveBook · Manning

Adversarial attack

As explained in Inside AI

An adversarial example is an input deliberately altered to cause a machine-learning model to produce an incorrect or unreliable output.

An adversarial attack deliberately alters an input so a machine-learning model produces an incorrect or unreliable result. The changes can be subtle enough to escape ordinary inspection while disrupting the model’s behavior. section 3.4.2

Surveillance-system disruption

Subtle pixel alterations can disrupt a deep-learning surveillance system designed to detect intruders. Creating such attacks may not be straightforward, but the vulnerability raises serious questions for mission-critical uses. section 3.4.2

Adversarial attacks matter because they expose limits in deployment:

  • They show that strong performance does not automatically make a deep-learning system trustworthy. section 3.4.2
  • They are especially concerning when errors affect safety or high-stakes decisions. section 3.4.2

How Inside AI covers it

How deep it goes
Defines subtle input alteration and illustrates its effect, but does not explain attack-generation methods or defenses.
Examples
Uses subtly modified pixels that disrupt an intruder-detection surveillance system.
How it connects
Links adversarial attacks with deep learning, surveillance systems, black-box models, and trustworthy AI.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Inside AI. The section links point to the parts of the book it draws on.