問題文
A security services firm's incident classification model is retrained monthly on the reports that supervisors have confirmed. An auditor considers how the arrangement could be degraded from inside. Which route is most specific to it?
選択肢
- A supervisor confirming reports in a way that suits their own site, so that the labels the next retraining learns from shift toward that supervisor's preference.
- A repeated query to the model with variations of the same report to learn how it classifies, since understanding the model's behavior would allow that supervisor to phrase future reports so that they receive the classification they want from the system.
- An outside party gaining access to the firm's report system and inserting false reports, since an attacker who can write to the source of the training data has the most direct means of influencing what the model learns during its next monthly retraining cycle.
- An unusual wording in a guard's report, so that the model fails to classify it and the incident is missed.