AAIA ISACA Advanced in AI Audit - Set 2 - Part 1

Correct Answer: Alignment with business impact assessment

Option c is correct because data drift occurs when real-world data characteristics change over time compared to the original training data, causing model performance degradation as patterns become outdated.

Option a is incorrect because data poisoning involves malicious injection of bad data into training sets, not natural changes in real-world data patterns over time.

Option d is incorrect because adversarial evasion involves crafted inputs designed to fool models, not gradual performance decline due to shifting real-world conditions.

Option b is incorrect because model extraction refers to adversaries stealing model logic through queries, which is unrelated to performance degradation from changing data patterns.

Correct Answer: Backing up trained model weights

Option a is correct because backing up trained model parameters and weights allows organizations to recreate exact model versions, enabling auditors to understand model state during specific decisions.

Option b is incorrect because data minimization focuses on reducing data collection to necessary amounts, which does not directly support model version reproducibility for audits.

Option c is incorrect because homomorphic encryption enables computation on encrypted data but does not address the ability to recreate specific model versions.

Option d is incorrect because data classification determines sensitivity levels for security purposes but does not enable recreation of exact model versions for auditing.

Correct Answer: Data integrity compromise

Option d is correct because data integrity ensures data is accurate, complete, and reliable. Systematic errors from malfunctioning equipment directly compromise training data accuracy and model reliability.

Option a is incorrect because prompt injection involves crafting malicious inputs to bypass safety rules, not systematic data quality issues from faulty collection equipment.

Option b is incorrect because configuration management tracks system components and versions, which is unrelated to data quality issues from malfunctioning collection equipment.

Option c is incorrect because adversarial inference involves attacks designed to extract information from models, not data quality problems from equipment malfunctions.

Correct Answer: Development demarcation practices

Option c is correct because development demarcation requires using non-production environments with synthetic data, preventing exposure of sensitive production data during development and testing activities.

Option a is incorrect because while proactive prevention is important, the specific violation here concerns mixing production data with development environments rather than general risk anticipation.

Option b is incorrect because full functionality refers to achieving both privacy and operational goals simultaneously, not the separation of development from production data.

Option d is incorrect because visibility and transparency relate to informing users about data handling, not the separation of development and production environments.

Correct Answer: Privacy as default setting

Option a is correct because privacy as default ensures personal data is automatically protected without requiring user action, meaning privacy-preserving options are pre-selected and data collection minimized.

Option b is incorrect because end-to-end security focuses on protection throughout the entire lifecycle rather than automatic privacy protection without user intervention.

Option c is incorrect because embedded design integration refers to building privacy into architecture fundamentally, not specifically automatic protection without user action.

Option d is incorrect because user-centric respect emphasizes user control and empowerment over data rather than automatic protection requiring no user action.

Correct Answer: Maintaining tokenization framework consistency

Option d is correct because the same tokenization framework, configuration, and version used during training must be maintained during production to prevent performance issues from text processing differences.

Option a is incorrect because data drift monitoring detects changes in real-world data patterns, not inconsistencies in text processing library configurations between training and production.

Option b is incorrect because adversarial testing evaluates resistance to malicious inputs, not compatibility issues arising from library updates affecting text processing methods.

Option c is incorrect because input validation ensures data quality and format correctness but does not address version consistency requirements for text processing components.

Correct Answer: Data poisoning attack

Option b is correct because data poisoning involves maliciously injecting bad or biased data into training sets, causing the model to learn incorrect patterns that attackers can later exploit.

Option a is incorrect because model extraction involves stealing model logic through systematic queries to replicate the model, not manipulating training data.

Option c is incorrect because adversarial evasion uses specially crafted inputs to fool models during inference, not attacks on the training data itself.

Option d is incorrect because prompt injection manipulates model inputs during operation to bypass safety rules, not the training data used to build the model.

Correct Answer: Black box decision processes

Option c is correct because deep learning models with millions of parameters process data through opaque internal mechanisms, making it extremely difficult to explain specific decisions clearly.

Option a is incorrect because data minimization concerns limiting data collection rather than understanding how models arrive at specific recommendations or decisions.

Option b is incorrect because configuration management tracks system components and versions but does not address the fundamental challenge of understanding model decision logic.

Option d is incorrect because human oversight refers to supervision capabilities rather than the inherent difficulty of interpreting complex neural network decision processes.

Correct Answer: Documenting bias risk considerations

Option d is correct because the Plan and Design stage requires identifying and documenting potential risks including ethical and bias concerns before development begins, as mistakes here ripple throughout.

Option a is incorrect because fairness testing protocols are implemented during the Test, Evaluate, and Validate stage, not during initial planning and design phases.

Option b is incorrect because monitoring dashboards are deployed during the Operate and Monitor stage after the AI system is live in production environments.

Option c is incorrect because stress testing occurs during later validation stages to evaluate performance under pressure, not during initial planning activities.

Correct Answer: Concept drift occurrence

Option c is correct because concept drift occurs when underlying relationships in the data fundamentally change over time, making previously learned patterns invalid even with similar data.

Option a is incorrect because data quality degradation refers to accuracy and completeness problems in data itself rather than changes in fundamental relationships between variables.

Option b is incorrect because statistical sampling drift relates to collection methodology issues rather than fundamental changes in how variables relate to each other.

Option d is incorrect because model extraction refers to adversarial theft of model logic through queries, not performance degradation from changing relationships in data.

Correct Answer: Rollback plan preparation

Option d is correct because rollback plans provide predefined procedures to quickly revert to a stable previous state if serious problems emerge after deployment to production environments.

Option a is incorrect because while user training is important for proper system use, it does not directly enable quick recovery when serious technical problems emerge.

Option b is incorrect because notification processes inform stakeholders about deployment status but do not provide mechanisms for rapid recovery from serious issues.

Option c is incorrect because installation documentation supports proper deployment execution but does not establish procedures for quick recovery from post-deployment problems.

Correct Answer: Systemic bias from data

Option a is correct because systemic bias stems from historical biases embedded in institutions and societal structures that are reflected in training data, causing AI to learn discriminatory patterns.

Option b is incorrect because statistical sampling error relates to unrepresentative data collection rather than historical discriminatory patterns embedded in institutional practices reflected in data.

Option c is incorrect because human labeling mistakes involve individual errors during data preparation rather than systematic historical discrimination patterns in institutional data.

Option d is incorrect because algorithmic configuration problems relate to technical settings rather than historical societal discrimination patterns captured in training data.

Correct Answer: Investigate performance metrics

Option d is correct because investigating performance metrics helps identify root causes of degradation before taking corrective action, ensuring responses are appropriate and targeted to actual problems.

Option a is incorrect because complete retraining is resource-intensive and premature without first understanding through metrics investigation what is actually causing the performance issues.

Option b is incorrect because increasing human oversight addresses symptoms rather than root causes and should follow investigation that identifies what specific issues require oversight.

Option c is incorrect because replacing systems entirely is drastic and should only occur after investigation determines the current system cannot be effectively remediated.

Correct Answer: Model backdoor insertion

Option c is correct because model backdoors are hidden functionalities intentionally inserted into models that activate on specific trigger inputs, representing a serious supply chain security threat.

Option a is incorrect because dependency library attacks involve compromised external code libraries rather than hidden functionality embedded directly within the pre-trained model itself.

Option b is incorrect because engineering environment compromise involves attacks on development infrastructure rather than malicious functionality hidden within vendor-provided models.

Option d is incorrect because training pipeline corruption refers to attacks on data processing workflows rather than intentionally hidden functionality within delivered model components.

Correct Answer: Initial Plan and Design stage

Option a is correct because Plan and Design is the most critical stage where risk considerations including safety must begin, as mistakes here ripple through everything and become harder to fix.

Option b is incorrect because the Build and Adapt stage focuses on technical model development, and safety considerations should already be documented from the planning phase.

Option c is incorrect because Operate and Monitor occurs after deployment when safety issues may already be embedded, making remediation significantly more difficult and costly.

Option d is incorrect because testing validates against requirements that should already incorporate safety considerations established during the planning phase.

Correct Answer: Logging key inference steps

Option d is correct because logging key model inferences, inputs, and intermediate steps creates visibility into AI decision processes, enabling investigation and understanding of unexpected results.

Option a is incorrect because access restrictions control who can interact with systems but do not provide visibility into why models produce specific outputs or decisions.

Option b is incorrect because benchmark testing compares performance against baselines but does not explain why specific individual results occur in operational contexts.

Option c is incorrect because backup procedures support recovery and reproducibility but do not provide insight into decision processes causing unexpected results.

Correct Answer: Containing autonomous actions

Option c is correct because during incident response, the first priority is containment to limit the AI's autonomous actions and increase human oversight, preventing further damage while investigation proceeds.

Option a is incorrect because root cause analysis is important but occurs after initial containment ensures the AI is not continuing to cause additional problems or damage.

Option b is incorrect because permanent fixes require understanding the root cause first and should not be rushed before properly containing the immediate incident.

Option d is incorrect because documenting lessons learned is part of post-incident review, which occurs after the incident has been contained and resolved.

Correct Answer: Data scientist training access

Option b is correct because AI environments require expanded access controls covering data scientists accessing training data, MLOps engineers deploying models, and other AI-specific roles beyond traditional users.

Option a is incorrect because network segmentation is a standard IT security control applied similarly to both traditional IT systems and AI environments.

Option c is incorrect because password complexity standards are fundamental IT security controls that apply equally to traditional IT and AI systems without AI-specific differences.

Option d is incorrect because firewall configuration is a standard network security control implemented similarly across traditional IT infrastructure and AI hosting environments.

Correct Answer: Sampling bias in collection

Option c is correct because sampling bias occurs when training data does not representatively cover all populations the AI will serve, causing poor performance for underrepresented groups.

Option a is incorrect because data integrity refers to accuracy, completeness, and reliability of data rather than representativeness of different customer populations in training sets.

Option b is incorrect because classification level errors involve incorrect sensitivity categorization of data rather than unrepresentative population coverage in training datasets.

Option d is incorrect because confidentiality control failures involve unauthorized data access rather than unrepresentative population sampling during data collection processes.

Correct Answer: Securing model weight disposal

Option d is correct because trained model weights represent significant intellectual property investment and must be securely disposed of to prevent unauthorized access or theft during decommissioning.

Option a is incorrect because stakeholder communications are important for coordination but do not directly protect intellectual property assets from unauthorized access during retirement.

Option b is incorrect because archiving compliance documentation supports regulatory requirements but does not address securing proprietary model assets from unauthorized access.

Option c is incorrect because migrating useful data supports business continuity but does not specifically address protecting intellectual property contained within trained model components.