Securing AI Systems
Free CompTIA SecAI+ (CY0-001) practice — 6 questions on Securing AI Systems, with explanations. No sign-up. Full 12-question mixed test →
During red team testing of a computer vision model used for automated loan approval, attackers successfully modified a single pixel in uploaded identity documents to cause 23% of fraudulent applications to be approved. The model uses ResNet-50 architecture with standard training. Which defense provides the MOST robust protection against this adversarial attack pattern?
Adversarial training explicitly teaches models to correctly classify inputs with adversarial perturbations, providing learned robustness. This is currently the most effective defense against known adversarial attack patterns and generalizes better than preprocessing defenses.
A healthcare AI system processes patient diagnostic images. Compliance requires a complete audit trail showing which training data influenced each diagnosis. The current model is a deep neural network with 89M parameters. Which approach BEST satisfies the auditability requirement without complete model redesign?
Influence functions mathematically trace how removing or modifying specific training examples would affect model predictions, providing the required training data provenance. This works with existing deep learning models without architectural changes.
An organization's AI development pipeline includes automated model training on cloud infrastructure. Security audit reveals that model checkpoints stored in object storage contain embedded API keys for the production database in the optimizer state. The keys were present in training data comments. Which control MOST effectively prevents credential leakage in future training runs?
Preventing secrets from entering the training pipeline at the source is the most effective control. Pre-training sanitization removes credentials before they can be memorized by the model, addressing the root cause rather than symptoms.
A company fine-tunes GPT-4 on proprietary customer service conversations. Post-deployment monitoring detects that specific prompt patterns can extract verbatim training examples including customer PII. The model must remain available while fixing this issue. Which remediation provides IMMEDIATE risk reduction without model downtime?
Prompt filtering at the API layer provides immediate protection without model changes or downtime. Blocking known extraction patterns (e.g., 'repeat after me', 'what did the customer say') prevents the attack vector while maintaining service availability.
A machine learning model for fraud detection shows 94% accuracy overall but only 67% accuracy for transactions from users in ZIP codes with median income below $35,000. Investigation reveals the training dataset had 10x fewer examples from these areas. Which technique BEST addresses this fairness issue while maintaining fraud detection effectiveness?
Equalized odds constraints ensure that true positive and false positive rates are similar across demographic groups, directly addressing the fairness metric disparity. Post-processing adjustments maintain the base model while enforcing fairness without requiring retraining or new data.
An autonomous vehicle ML system uses sensor fusion from cameras, lidar, and radar. Security testing reveals that projecting specific infrared patterns onto road signs causes the system to misclassify stop signs as speed limit signs in 34% of trials. Which mitigation MOST effectively defends against this multi-modal adversarial attack?
Multi-modal sensor fusion's key advantage is redundancy; if infrared projections fool the camera but lidar/radar correctly identify the stop sign geometry, divergence detection provides robust defense. This leverages the fundamental property that adversarial attacks typically don't transfer across different sensor physics.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$109.99 $34.99 with code FREETEST33 — valid through August 23.
Get my $34.99 deal →