Which security concern is most closely associated with the Training stage of the AI lifecycle?

• A. Data poisoning
• B. Data integrity and access control
• C. Model exposure
• D. Inference security

Answer: (B)

Training relies on trusted data. Ensuring data integrity means the training data hasn’t been altered, mislabeled, or tampered with in any way, so the model learns correct patterns. Access control ensures only authorized people can read or modify that data, preventing unauthorized changes or leakage of sensitive training samples. Together, these protections keep the learning process trustworthy and secure, reducing the risk of the model being trained on poisoned or compromised data and safeguarding the data used to teach it.

Data poisoning is a direct way to disrupt training by injecting malicious examples, but it’s mitigated by strong data integrity checks and strict access controls—the broader, foundational protections during training. Model exposure and inference security are more about protecting the model or its outputs after training, rather than the training process itself.