Adversarial ML / poisoning
Poisoning
If an attacker controls part of the training corpus, they can install behaviour that lies dormant until a trigger appears. Scraped pre-training data makes the threat model uncomfortably realistic.
1 entry.
Adversarial ML / poisoning
If an attacker controls part of the training corpus, they can install behaviour that lies dormant until a trigger appears. Scraped pre-training data makes the threat model uncomfortably realistic.
1 entry.