Data Poisoning
Data Poisoning
A cyberattack that corrupts an AI's training data to manipulate its decisions.
In Simple Terms
Data Poisoning is a cyberattack technique that feeds an AI false data ahead of time, so it later makes the wrong call exactly as the attacker intends. It exploits the way AI builds its judgment rules from data, mixing in corrupted data to paralyze the AI's ability to tell things apart. It's used for things like helping spam emails slip past automatic filters or disabling security checks.
Behind the Name
The name "Data Poisoning" combines "data" with "poisoning," meaning to introduce poison into something. It's called that because the attack works by slipping harmful information into an AI's training data, corrupting it from the inside — just like poison spreading through something.
Take a Closer Look!
Data Poisoning is a security attack that mixes corrupted information into an AI's training data to trigger faulty behavior.
An AI learns rules and patterns by reading through huge amounts of data, and this attack works by planting traps in that very data it depends on.
Put simply, it's like secretly writing wrong answers into the textbook an AI studies from, so it gives incorrect answers on the test later.
Before the AI is trained, attackers deliberately mix in data with false labels or data laced with special noise.
As a result, the finished AI ends up misbehaving on purpose only when certain conditions are met, or overlooking dangerous data it should have caught.
Defending against this attack requires carefully vetting where training data comes from and checking it in advance for anything abnormal.
It's also important to keep monitoring how the AI behaves when it makes predictions, watching for an increase in unnatural judgments, to keep things safe.