Skip to content

Privacy & Security in AI

Protecting data and defending models — training on sensitive data safely, and keeping AI systems robust against attack.

Modern AI learns by studying huge piles of information — sometimes information about real people, like their messages, photos, or medical records. That creates two big worries. Privacy: the AI must give helpful answers without ever blurting out one person's private details. Security: clever attackers may try to trick or corrupt it.

Imagine a friend who has secretly read everyone's diaries so they can give great advice. You want their advice — but you never want them to quote your diary word-for-word to a stranger. And you'd worry if someone slipped a fake diary into the pile to change what your friend believes, or asked sneaky questions to fish out your secrets. Keeping that friend both helpful and trustworthy is exactly what this field is about.

The main ideas

  • Federated learning — Train across many devices without centralizing raw data.
  • Differential privacy — Provable guarantees that a model doesn't leak any individual's data.
  • Adversarial examples — Tiny crafted perturbations that fool models — and defenses against them.
  • Data poisoning & backdoors — Corrupting training data to plant hidden behaviors.
  • Model extraction & inversion — Stealing a model or reconstructing its training data through its API.
  • Prompt injection & LLM security — The new attack surface of agents and RAG — untrusted input hijacking instructions.

AI Safety, Alignment & Ethics · AI Agents & Autonomy · AI Ethics & Governance

Want to make things?

Head to AI School — AI camps where kids build their own games.