Transformer Security & AI Vulnerabilities Quiz

1. What is the primary security concern with fine-tuning pre-trained transformers?

Fine-tuning always improves model accuracy

The base model may retain vulnerabilities or biases from pretraining that persist after fine-tuning

Fine-tuning requires removing all attention mechanisms

Fine-tuning is computationally impossible

2. What is the primary benefit of using 'secure multi-party computation' (MPC) for transformer inference?

It increases inference speed significantly

It allows multiple parties to compute on sensitive data without revealing individual inputs

It removes the need for model validation

It eliminates all attention mechanisms

3. How can 'knowledge distillation' introduce security risks in transformer systems?

It always improves model robustness

A smaller student model trained on teacher outputs may leak proprietary knowledge more easily

It removes all model parameters

It eliminates the need for training data

4. Which security practice is essential for transformer model deployment in production?

Continuous monitoring, logging, and anomaly detection for model behavior

Running inference without any authentication

Storing model weights in plain text

Disabling all input preprocessing

5. What is 'poisoning attack' in the context of transformer training data?

Intentionally injecting malicious or incorrect data to corrupt model behavior

Removing sensitive data from training sets

Increasing model validation accuracy

Using only certified training datasets

6. True or False: Larger transformer models are automatically more secure than smaller models.

True

False

7. How does model watermarking protect transformer intellectual property?

It encrypts all model parameters

It embeds hidden patterns or signatures to prove ownership and detect unauthorized copies

It removes attention mechanisms from the architecture

It prevents any model inference

8. What role does secure enclave technology (e.g., Intel SGX) play in transformer security?

It has no relevance to AI security

It can protect model inference and sensitive computations from unauthorized access

It increases model training speed

It removes the need for input validation

9. Which of the following best describes a 'side-channel attack' on transformer deployments?

Exploiting timing, power consumption, or cache behavior to extract model information

Directly modifying model weights on disk

Using only open-source transformer implementations

Increasing batch size during inference

10. True or False: Homomorphic encryption can allow secure inference on transformers without decrypting input data.

True

False

11. What is the primary security vulnerability in transformer attention mechanisms that could allow unauthorized information leakage?

Attention heads can reveal patterns about sensitive training data through gradient analysis

Positional encodings are always deterministic and predictable

Layer normalization removes all security protections

Query-key similarity scores are immune to extraction attacks

12. How can adversarial training improve transformer security?

By training on deliberately crafted adversarial examples to improve robustness

By removing all regularization techniques

By increasing model parameters exponentially

By using only synthetic data

13. What is a membership inference attack against transformer models?

An attempt to determine if a specific data point was in the training set

Replacing model weights with random values

Blocking user access to the model API

Modifying attention head dimensions

14. Which security measure is most critical for protecting proprietary transformer models from unauthorized access?

Model quantization to reduce file size

Access controls, encryption at rest, and secure API authentication

Removing all documentation from the codebase

Training only on public datasets

15. True or False: Transformers are inherently resistant to adversarial examples because of their attention mechanisms.

True

False

16. What does 'federated learning' help protect in distributed transformer training?

Model inference speed only

User privacy by keeping training data local rather than centralizing it

Model parameter size

Attention head complexity

17. Which of the following is a backdoor attack on transformers?

Injecting hidden triggers during training to cause specific malicious outputs

Reducing model accuracy through quantization

Increasing inference latency intentionally

Adding more attention heads to the model

18. How does differential privacy protect transformer models during training?

It encrypts all model parameters with RSA

It adds calibrated noise to gradients to prevent memorization of individual training examples

It removes all personal information from input data

It stops model training after a fixed number of steps

19. What is a model extraction attack in the context of transformer security?

Physically removing model weights from servers

Querying a model repeatedly to recreate its behavior and steal intellectual property

Deleting transformer layers during inference

Removing attention heads from the architecture

20. Which technique best mitigates prompt injection attacks in transformer-based language models?

Increasing model parameters to improve robustness

Input validation, sanitization, and role-based prompting with clear delimiters

Disabling all user input features

Using only smaller, less capable models

SecAI+ Transformer Architecture Security Considerations Quiz

1. What is the primary security concern with fine-tuning pre-trained transformers?

2.

What first name or nickname would you like us to use?

2. What is the primary benefit of using 'secure multi-party computation' (MPC) for transformer inference?

3. How can 'knowledge distillation' introduce security risks in transformer systems?

4. Which security practice is essential for transformer model deployment in production?

5. What is 'poisoning attack' in the context of transformer training data?

6. True or False: Larger transformer models are automatically more secure than smaller models.

7. How does model watermarking protect transformer intellectual property?

8. What role does secure enclave technology (e.g., Intel SGX) play in transformer security?

9. Which of the following best describes a 'side-channel attack' on transformer deployments?

10. True or False: Homomorphic encryption can allow secure inference on transformers without decrypting input data.

11. What is the primary security vulnerability in transformer attention mechanisms that could allow unauthorized information leakage?

12. How can adversarial training improve transformer security?

13. What is a membership inference attack against transformer models?

14. Which security measure is most critical for protecting proprietary transformer models from unauthorized access?

15. True or False: Transformers are inherently resistant to adversarial examples because of their attention mechanisms.

16. What does 'federated learning' help protect in distributed transformer training?

17. Which of the following is a backdoor attack on transformers?

18. How does differential privacy protect transformer models during training?

19. What is a model extraction attack in the context of transformer security?

20. Which technique best mitigates prompt injection attacks in transformer-based language models?