1. Which core architecture, introduced in 2017, relies entirely on self-attention mechanisms and forms the foundation of modern Large Language Models (LLMs) like GPT-4?
A) Convolutional Neural Network (CNN)
B) Recurrent Neural Network (RNN)
C) Transformer Architecture
D) Generative Adversarial Network (GAN)
Answer: C) Transformer Architecture
2. In the context of Advanced AI and cybersecurity, what is “Prompt Injection”?
A) A technique to speed up model inference times by injecting code directly into GPU memory.
B) A vulnerability where an attacker manipulates the input of a language model to bypass safety filters or execute unauthorized instructions.
C) A method of securely transferring training weights between neural networks.
D) An automated tool used to generate synthetic training text for cybersecurity datasets.
Answer: B) A vulnerability where an attacker manipulates the input of a language model to bypass safety filters or execute unauthorized instructions.
3. Which advanced machine learning paradigm allows an agent to learn optimal behaviors by trial and error through rewards and penalties from its environment?
A) Supervised Learning
B) Unsupervised Learning
C) Reinforcement Learning
D) Transfer Learning
Answer: C) Reinforcement Learning
4. What does “RLHF” stand for in modern LLM development and alignment?
A) Recursive Learning with Heuristic Feedback
B) Reinforcement Learning from Human Feedback
C) Rapid Layer-wise Hyperparameter Filtering
D) Robust Language Harmonization Framework
Answer: B) Reinforcement Learning from Human Feedback
5. Which of the following generative AI models typically consists of a “Generator” and a “Discriminator” competing against each other in a zero-sum game?
A) Variational Autoencoders (VAEs)
B) Generative Adversarial Networks (GANs)
C) Diffusion Models
D) Autoregressive Transformers
Answer: B) Generative Adversarial Networks (GANs)
6. What is “Model Quantization” primarily used for in advanced deep learning deployment?
A) Converting continuous text tokens into numerical vector embeddings.
B) Increasing the number of parameters to improve generative accuracy.
C) Reducing the memory footprint and computational cost of a model by lowering the precision of its weights (e.g., from FP32 to INT8).
D) Automatically splitting a neural network across multiple cloud servers.
Answer: C) Reducing the memory footprint and computational cost of a model by lowering the precision of its weights (e.g., from FP32 to INT8).
7. In computer vision and multimodal AI, what is a “Zero-Shot Learning” capability?
A) The ability of a model to perform a task or recognize objects it has not been explicitly trained or fine-tuned on.
B) Training a neural network using zero training datasets to save storage.
C) The process of completely erasing historical chat context to protect user privacy.
D) An optimization technique that reduces training time to zero seconds.
Answer: A) The ability of a model to perform a task or recognize objects it has not been explicitly trained or fine-tuned on.
8. Which advanced threat vector involves feeding subtly modified input data to an AI model to cause it to misclassify objects or output erroneous decisions (e.g., fooling an autonomous car’s vision system)?
A) Data Poisoning Attack
B) Adversarial Attack
C) Model Inversion Attack
D) Denial-of-Service Attack
Answer: B) Adversarial Attack
9. What is the primary purpose of Retrieval-Augmented Generation (RAG) in AI architectures?
A) To encrypt prompts before they reach the language model server.
B) To allow an LLM to fetch external, up-to-date, or private data from a knowledge base to ground its responses and reduce hallucinations.
C) To compress large text datasets into smaller vector files.
D) To accelerate hardware rendering speeds for real-time video generation.
Answer: B) To allow an LLM to fetch external, up-to-date, or private data from a knowledge base to ground its responses and reduce hallucinations.
10. What is “Model Inversion” in the context of AI privacy and security?
A) Reversing the order of layers in a deep neural network to test stability.
B) An attack where an adversary reconstructs private training data (such as facial features or medical records) by querying the model’s outputs.
C) Switching a model from supervised learning to unsupervised learning post-deployment.
D) A technique used to downgrade a newer AI model to an older software version.
Answer: B) An attack where an adversary reconstructs private training data (such as facial features or medical records) by querying the model’s outputs.
