Ethical, Societal, and Security Concerns in AI AIE Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A city government is considering an AI system to screen job applicants for public sector positions. The system uses historical hiring data, which underrepresents certain minority groups. Civil rights groups are concerned about inequality and lack of transparency. The city wants to deploy the system but must address these concerns. What should the city do first?

   • AImmediately discard the AI system and rely on human recruiters only.
   • BConduct a bias audit on the training data and model outcomes, and publish a transparency report.
   • CAdd more training data from underrepresented groups without further analysis.
   • DDeploy the system but keep the decision logic confidential to protect proprietary algorithms.
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   A defense contractor develops an AI system for drone navigation. The system relies on GPS and visual landmarks. The contractor is concerned about adversarial attacks that could cause the drone to crash. The contractor has a limited budget and must choose between two mitigations: (1) adversarial training to make the model robust to image perturbations, or (2) adding a redundant sensor system (e.g., inertial navigation) that does not rely on the AI model. Which approach should the contractor prioritize to maximize safety?

   • AAdversarial training, because it directly strengthens the AI model against the most likely attack vector
   • BBoth, but start with adversarial training because it is cheaper and faster to implement
   • CRedundant sensor system, because it provides a fallback that does not depend on the AI model's vulnerabilities
   • DNeither, because the drone's safety is already ensured by the AI model's high accuracy
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A healthcare startup uses an AI model to predict patient readmission risk. The model is hosted on a cloud API and is used by multiple hospitals. A security researcher warns that the model could be stolen by sending many queries and observing the outputs. The startup wants to protect the model without degrading service for legitimate users. Which mitigation should they implement?

   • AMake the model's predictions less accurate to deter attackers.
   • BRemove the API and share the model weights only with trusted hospitals.
   • CAdd a watermark to the model's training data so stolen models can be identified.
   • DImplement rate limiting and monitor for unusual query patterns.
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   A self-driving car company uses a deep learning model to detect pedestrians. During testing, the team discovers that placing a small sticker on a stop sign causes the model to classify the sign as a speed limit sign. The company wants to reduce this vulnerability before the cars are deployed on public roads. Which approach is most effective?

   • AUsing a larger neural network with more parameters
   • BDisabling the model's ability to classify speed limit signs
   • CAdding more training images of stop signs in different lighting conditions
   • DAdversarial training by augmenting the training set with perturbed images
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A cybersecurity firm develops an AI model to detect malware. The model is trained on a public dataset of malware samples. An attacker poisons the dataset by injecting benign files labeled as malware, causing the model to misclassify benign files as malicious. The firm wants to prevent this. What should they do?

   • AIncrease the model's sensitivity to detect more malware.
   • BRegularly retrain the model with the same dataset.
   • CValidate the integrity of training data sources and use anomaly detection to filter outliers.
   • DUse a more complex model to better fit the training data.
   Reveal Solution Discussion Select an answer first