CY0 CompTIA SecAI+ Beta Free Practice Test - 20 Questions
This practice bank exercises knowledge of AI security threats and defenses, including adversarial examples, data poisoning, model poisoning, model inversion, membership inference, and model stealing. It covers privacy-preserving techniques (differential privacy, federated learning, homomorphic encryption), explainable AI, bias mitigation, secure deployment (TEEs, input validation), and model provenance. Each question targets a specific objective, such as identifying attack types, applying defenses, or evaluating trade-offs. Mastery of these topics is essential for understanding how to secure AI systems throughout their lifecycle.
What this CY0 practice set measures
This is an analysis of the practice bank, not a claim about the vendor's live exam blueprint. Use it to identify the knowledge, judgment, and recall patterns exercised here, then verify your coverage against the current official exam guide.
1.1 Identify Adversarial Attacks on AI Systems
Adversarial attacks manipulate inputs to cause misclassification. An adversarial example adds small, imperceptible perturbations to data, while an adversarial patch is a physical object like a sticker that induces errors. These attacks exploit model sensitivity and are distinct from data poisoning (training data corruption). Understanding these vectors is critical for defending against inference-time threats.
- Adversarial examples use subtle input perturbations to trigger misclassification.
- Adversarial patches are physical objects that cause model errors when present in the scene.
- Both attacks target model prediction rather than training data.
1.2 and 1.4 Describe Privacy Attacks and Model Poisoning
Privacy attacks extract sensitive information from models. Model inversion reconstructs training data, membership inference determines if a record was used, and model stealing replicates functionality. In federated learning, model poisoning corrupts local updates to manipulate the global model. These attacks highlight the need for data protection and robust aggregation.
- Model inversion reconstructs representative samples from training data.
- Membership inference checks if a specific data point was in training.
- Model stealing copies model behavior via querying.
- Model poisoning attacks target federated learning by altering local model updates.
2.1, 2.2, 2.3 Privacy Preservation and Collaborative Learning
Differential privacy adds noise to training to prevent membership inference. Federated learning with secure aggregation allows collaborative training without sharing raw data. Homomorphic encryption enables computation on encrypted data but incurs high computational overhead. Choosing the right technique balances privacy, functionality, and performance.
- Differential privacy guarantees that model outputs do not reveal individual record membership.
- Federated learning trains models locally; secure aggregation encrypts updates.
- Homomorphic encryption allows inference on encrypted data but is computationally expensive.
3.1, 3.2, 3.3 Explainability, Fairness, and Data Quality
Explainable AI provides transparency into model decisions, essential for defending against discrimination claims. Bias testing and mitigation address disparate impact. Data sanitization removes offensive or biased content to prevent harmful patterns. These practices ensure ethical and robust AI deployment.
- Explainable AI (XAI) helps interpret model decisions and defend against bias allegations.
- Bias testing and mitigation techniques reduce unfair outcomes across demographic groups.
- Data sanitization filters inappropriate content to avoid learning harmful biases.
Practice CY0 CompTIA SecAI+ Beta with real flashcards
Read the prompt, commit to an answer, then flip the card. Move through the deck at your own pace and repeat any topic that does not come back quickly.
Card 1 of 20
1 reviewed this session
Static practice bank
Start the 20-question diagnostic
The complete question bank is embedded in this pre-rendered page. There is no database request or second content download when you begin.
A security analyst at a healthcare company is testing an image classification model that diagnoses skin conditions. The model misclassifies a benign mole as malignant after a small, imperceptible change to the image pixels. Which type of attack has occurred?
Show hint
1.1 Identify adversarial attacks on AI systems
Study workflow
Turn one CY0 attempt into a study plan
- 1
Validate and Sanitize Training Data
Before training, inspect data for biases, offensive language, or potential poisoning triggers. Remove or appropriately handle harmful samples to prevent the model from learning undesirable patterns. Use automated filtering and manual review to ensure data integrity.
- 2
Implement Differential Privacy During Training
Add calibrated noise to the training process or gradients to bound the model's sensitivity to individual records. This protects against membership inference even if an attacker gains access to the model. Choose privacy budget (ε) appropriate for the use case.
- 3
Use Federated Learning with Secure Aggregation
When training across multiple data owners, keep data local and only share encrypted model updates. Secure aggregation combines these updates without revealing individual contributions, preventing data leakage and model poisoning.
- 4
Deploy Models in Trusted Execution Environments (TEEs)
Run model inference inside a hardware-enclave like Intel SGX to isolate computations from the host OS. This protects model parameters and user data from unauthorized access, even if the system is compromised.
- 5
Apply Input Validation and Sanitization at Inference
Validate all inputs against expected schemas, ranges, and types. Use sanitization libraries to strip potentially malicious content (e.g., SQL injection, prompt injection). This prevents attackers from manipulating model behavior through crafted inputs.
FAQ
Questions about this CY0 practice page
Clear boundaries on what the bank covers, how to use it, and where official vendor information still matters.
What is the difference between data poisoning and model poisoning?+
Data poisoning corrupts the training data itself, causing the model to learn incorrect associations. Model poisoning specifically targets federated learning by sending malicious local model updates to manipulate the global model. Both aim to degrade model performance but differ in the attack vector.
How does differential privacy protect against membership inference?+
Differential privacy adds noise to training algorithms so that the model's outputs are approximately the same whether any individual record is included or not. This makes it infeasible for an attacker to determine if a specific person's data was used, preserving membership privacy.
What is a trusted execution environment (TEE) and why is it important for AI security?+
A TEE is a hardware-isolated environment (e.g., Intel SGX) that protects code and data during execution. It ensures that even if the operating system is compromised, the model and its inputs remain confidential and tamper-free, crucial for sensitive inference tasks.
What are the trade-offs of using homomorphic encryption for inference?+
Homomorphic encryption allows computation on encrypted data, preserving privacy. However, it imposes significant computational overhead, making it impractical for real-time applications. It also does not protect against adversarial examples or model theft on its own.
Why is explainable AI (XAI) important for regulatory compliance?+
XAI provides transparency into how models reach decisions. In regulated sectors like finance and healthcare, this is essential for demonstrating fairness, avoiding discrimination claims, and meeting legal requirements to explain adverse outcomes.
Build the next review session
Browse another free bank or use the study strategy guide to turn your misses into spaced review.