What's changed: Created Professional Machine Learning Engineer Chapter 6 (Domain 6 "Monitoring": secure AI systems (theft/poisoning defense), Google Responsible AI (bias monitoring), fairness/readiness assessment, Vertex Explainable AI (feature attributions); Model Monitoring on Gemini Enterprise Agent Platform continuous evaluation, training/serving skew, feature attribution drift, performance vs baselines/simpler models/over time, common training/serving error monitoring).
6.1AI risks and Responsible AI
Understand building secure AI systems by protecting against unintentional exploitation of data/models (hacking), aligning with Google's Responsible AI practices (monitoring for bias), assessing AI solution readiness (fairness, bias), and model explainability on Agent Platform (Explainable AI).
AI brings risks alongside benefits. Build trustworthy AI by designing in security, fairness, and explainability.
6.1.1Secure AI and Responsible AI
Protect AI systems from unintentional exploitation (attacks) on training data/models—guard against theft/poisoning (data poisoning) and adversarial inputs with least access, VPC Service Controls, CMEK, and input validation. Align with Google's Responsible AI practices: build in bias monitoring, fairness, and suppression of harmful outputs. Before release, assess AI solution readiness for fairness, bias, and safety. Map "protect training data/models = least access + perimeter + encryption" and "unbiased AI = bias monitoring and fairness evaluation."
6.1.2Explainability (Explainable AI)
To surface a model's rationale, use Vertex Explainable AI. Attach feature attributions (how much each feature contributed to a prediction) to Vertex AI Prediction outputs, explaining why a prediction was made. This supports compliance, debugging, bias detection, and accountability to users. Tabular uses Shapley-value-based methods; images can use Integrated Gradients, etc. Map "explain a prediction's rationale per feature = Explainable AI (feature attributions)."
Common: requirement → means. E.g., "explain a prediction's rationale per feature" = Vertex Explainable AI (feature attributions); "detect/correct model bias" = bias monitoring/fairness evaluation (Responsible AI); "prevent attacks on training data/models" = least access + VPC Service Controls + CMEK; "pre-release validation" = AI readiness (fairness/bias/safety) assessment.
Watch the mix-ups: (1) Explainable AI (why this prediction) vs Model Monitoring (detect prod degradation/skew) have different purposes. (2) Fairness/bias often stems from data—consider it from collection/preprocessing. (3) High accuracy but unfair/unexplainable is a deployment risk—meet Responsible AI.
6.1.3Section summary
- Secure AI = guard against theft/poisoning = least access + VPC Service Controls + CMEK + input validation
- Responsible AI = bias monitoring/fairness evaluation; assess readiness before release
- Explain prediction rationale = Vertex Explainable AI (feature attributions)
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. To explain a prediction by showing how much each feature contributed, which is best?
Q2. To continuously evaluate whether a model makes unfair predictions against certain attribute groups, which is best?
Q3. To protect an AI system from theft/poisoning (data poisoning) of training data/models, which combination is best?
Q4. Before release, to validate fairness, bias, and safety—what assessment is this?
Q5. Which correctly distinguishes Explainable AI from Model Monitoring?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.