FRM Part II · FRM Exam Part II · The Financial Stability Implications of Artificial Intelligence
A bank deploys a machine learning fraud-detection model. Attackers learn to probe it and craft transactions that slip just under its detection thresholds, and they also try to poison its retraining data. Which pairing correctly names these two threats and an appropriate response?
The threats are adversarial evasion, where inputs are crafted to avoid detection, and data poisoning, where training data is corrupted. Appropriate responses are strong model monitoring, input validation, secured retraining data pipelines and human oversight. Liquidity, hedging or variable-count fixes do not address deliberate manipulation.
- AAdversarial evasion and data poisoning; apply robust model monitoring, input validation, controlled retraining data pipelines and human oversightCorrect
- BModel drift and survivorship bias; increase the number of input variables
- CLiquidity mismatch and fire sales; raise HQLA holdings
- DBasis risk and convexity risk; rebalance the hedge portfolio
Explanation
Crafting inputs to avoid detection is adversarial evasion, and corrupting training data is data poisoning. Mitigants include monitoring for anomalies, validating inputs, securing and curating retraining data, and keeping human oversight. The other pairs describe unrelated statistical, liquidity or hedging issues.
Did you get it right without looking?
One question tells you little. A timed set on The Financial Stability Implications of Artificial Intelligence shows your real accuracy, how long you take and where you lose marks.
More The Financial Stability Implications of Artificial Intelligence questions
- A supervisor assesses systemic implications of AI-related cyber risk. Which scenario best illustrates how AI could amplify financial stabili…
- A risk manager at an asset manager reviews an AI trading model whose training data covers only 2012-2021, a period of low inflation and low …
- A bank's model risk team is told that its generative AI credit-assessment tool is hard to explain and relies on opaque training data. Which …
- A risk manager assesses a vendor-supplied AI model for fraud detection that is also used by many other banks, and the vendor relies on a sin…
- A bank deploys an AI fraud-detection model. Attackers begin probing it with crafted inputs and also corrupt part of the data used for period…
- A risk manager models a portfolio of two strategies, each with 10% annual volatility and equal weights. In normal conditions the correlation…