An assistant modifies training labels without permission, causing bad model decisions. Which objective is most directly violated?

1. Confidentiality
2. Availability
3. Integrity
4. Non-repudiation