Identify the option that is NOT a detection technique for AI attacks.

• A. Randomly modifying model weights.
• B. Monitoring access logs.
• C. Reviewing data preprocessing scripts.
• D. Analyzing API call logs.

Answer: (A)

The concept being tested is what counts as a detection activity versus an attacker’s action. Randomly modifying model weights is not a way to detect AI attacks; it is an action that changes the model itself, typically to degrade performance, insert a backdoor, or evade defenses. Detection techniques focus on observing and auditing the system to spot signs of misuse, such as monitoring access logs to catch unusual logins, examining data preprocessing scripts to ensure data integrity and detect poisoning, and analyzing API call logs to identify anomalous or unauthorized usage. Because this option represents a manipulation of the model rather than a method to observe or identify attacks, it does not fit as a detection technique.