GCP Generative AI Leader Practice Question 79: Data Poisoning ML Lifecycle Security
Google Cloud Generative AI Leader · Test 8 · question 9 of 10
A manufacturer is training a custom model on data submitted by its network of parts suppliers. The security team warns that if an attacker managed to slip corrupted or mislabeled records into that supplier data before training begins, the resulting model could learn incorrect patterns or a hidden backdoor, and the damage would be very hard to detect once the model is in production.
Which security risk is the team describing, and at what point in the ML lifecycle must it be addressed?
Show answer and explanation
Answer:
Data poisoning is when an attacker deliberately corrupts or manipulates training data so the resulting model learns incorrect behavior or a hidden vulnerability. Because it happens before and during training, the defense has to be built into the data ingestion and preparation stages, through validation, provenance checks, and access controls on submitted data, not fixed once the model is already deployed. This is why Google's security guidance treats security as a concern across the whole ML lifecycle rather than a step added at the end.
B is wrong because prompt injection is an attack on a live model through crafted user input at inference time, a different risk from corrupting the training data itself. C is wrong because model exfiltration is about someone stealing the trained model artifact, not about corrupting the data that trains it. D is wrong because denial of service targets the availability of a deployed endpoint, not the integrity of the training data described in the scenario.
Reference: Google's Secure AI Framework (SAIF)