What Does This Mean?
An attack on a machine-learning system in which the training data is deliberately contaminated so that the model makes more errors or learns a hidden behavior. A backdoor attack is a specific form in which the model behaves a chosen way when an input contains a particular trigger.
REAL-WORLD EXAMPLE
""The paper shows how a small share of poisoned documents can plant a backdoor in the model.""
LORE & ORIGIN
Wikipedia's adversarial machine learning article defines poisoning as contaminating the training dataset with data designed to increase errors in the output, and describes a backdoor attack as a particular case that teaches a specific behavior for inputs with a trigger; it cites a 2025 study finding that about 250 poisoned documents could introduce a denial-of-service backdoor across the model sizes tested. This is a security-research term; the entry describes the concept only and gives no method. Agent-era variants have their own entries.
MENTIONS OVER TIME
Radar observation data not yet available for this term.
RELATED SEARCHES
TOP PLATFORMS
Not enough Radar data yet to show a platform breakdown.
WHEN DID YOU FIRST HEAR THIS?
CLICK AN OPTION BELOW TO CAST YOUR VOTE.
[ VOTE TO REVEAL COMMUNITY RESULTS ]