Data poisoning
Corrupting training data to manipulate a model.
An attacker inserts or alters records in training data so the model learns harmful behaviour, such as ignoring a certain kind of fraud. Controls include data provenance, access control over training data, validation and anomaly detection.
On the exam
Manipulated training data is poisoning; manipulated inputs at run time are evasion or prompt injection.
Related terms
- Prompt injectionText that tricks a language model into ignoring its instructions.
- Data lineage and provenanceWhere data came from and what happened to it on the way.
Test yourself with today’s question of the day.