Pass the SISA Cyber Security for AI CSPAI Questions and answers with ExamsMirror

For continuous improvement in open-source LLM-based virtual assistants, RLHF integrates human evaluations to align model outputs with preferences, iteratively refining behavior without full retraining. This method uses reward models trained on feedback to guide policy optimization, enhancing interaction quality over time. It addresses limitations like initial biases or suboptimal responses by leveraging real-world user inputs, making the system adaptive and efficient. Unlike full retraining, RLHF is parameter-efficient and scalable, ideal for production environments. Security benefits include monitoring feedback for adversarial attempts. Exact extract: "Implementing RLHF allows continuous refinement of the assistant's interactions based on user feedback, avoiding the need for constant full retraining while improving performance." (Reference: Cyber Security for AI by SISA Study Guide, Section on AI Improvement Techniques in SDLC, Page 85-88).

LLM plugin compromises occur when extensions or integrations, like API-connected tools in systems such as ChatGPT plugins, are exploited, leading to unauthorized data access or injection attacks. Attackers might hijack plugins to leak user queries, training data, or system prompts, breaching privacy and enabling further escalations like lateral movement in networks. This risk is amplified in open ecosystems where plugins handle sensitive operations, necessitating vetting, sandboxing, and encryption. Unlike benefits like accuracy gains, compromises erode trust and invite regulatory penalties. Mitigation strategies include regular vulnerability scans, least-privilege access, and monitoring for anomalous plugin behavior. In AI security, this highlights the need for robust plugin architectures to prevent cascade failures. Exact extract: "A potential risk of LLM plugin compromise is unauthorized access to sensitive information, which can lead to data breaches and privacy violations." (Reference: Cyber Security for AI by SISA Study Guide, Section on Plugin Security in LLMs, Page 155-158).

Embedding an LLM 'as-is' means direct integration of the pretrained model into the app framework without alterations, relying on its inherent capabilities for tasks like text generation, simplifying SDLC by avoiding customization overhead. This is suitable for general-purpose apps but may lack optimization for specifics, contrasting with tailored approaches. It accelerates deployment while posing risks like unmitigated biases, necessitating post-integration safeguards. Exact extract: "It describes integrating the LLM without modifications, using out-of-the-box capabilities directly in the application." (Reference: Cyber Security for AI by SISA Study Guide, Section on LLM Integration Methods, Page 110-113).

ISO 42001 mandates a systematic risk treatment process, involving identification of AI risks (e.g., bias, security), analysis of impacts, evaluation against criteria, and development of treatment plans like mitigation or acceptance. This ensures proactive management throughout the AI lifecycle. Exact extract: "ISO 42001 requires identifying, analyzing, and evaluating AI risks with appropriate treatment plans." (Reference: Cyber Security for AI by SISA Study Guide, Section on Risk Treatment in ISO 42001, Page 270-273).

Transformer models leverage self-attention to process entire sequences concurrently, unlike RNNs, which handle inputs sequentially and struggle with long-range dependencies due to vanishing gradients. By computing attention scores across all words, Transformers capture both local and global contexts, enabling better modeling of relationships in tasks like translation or summarization. For example, in a long sentence, attention links distant pronouns to their subjects, improving coherence. This contrasts with RNNs’ sequential limitations, which hinder capturing far-apart dependencies. While parallelism (option C) aids efficiency, the core improvement lies in dependency modeling, not just speed. Exact extract: "The attention mechanism enables Transformers to attend to nearby and distant words simultaneously, significantly improving long-term dependency understanding over RNNs." (Reference: Cyber Security for AI by SISA Study Guide, Section on Transformer vs. RNN Architectures, Page 50-53).

Parameter-efficient fine-tuning (PEFT) strategies, like LoRA or adapters, freeze most pretrained parameters and train only lightweight modules, reducing computational costs while adapting to new tasks. This preserves general knowledge, prevents catastrophic forgetting, and enables quick deployments in resource-constrained settings. For LLMs, it's crucial for efficiency in production, allowing specialization without retraining billions of parameters. Security-wise, it minimizes exposure to new data risks. Exact extract: "A common strategy is freezing the majority of model parameters and updating only a small task-relevant subset, ensuring efficiency in fine-tuning for production deployment." (Reference: Cyber Security for AI by SISA Study Guide, Section on Efficient Fine-Tuning in SDLC, Page 90-92).

The transition from rigid rule-based systems, which rely on predefined logic and struggle with variability, to machine learning models introduced data-driven learning, allowing systems to adapt dynamically to new patterns with less human oversight. This shift boosted efficiency in automated tasks by enabling real-time adjustments, such as in spam detection where ML models evolve with threats, unlike static rules. It minimized manual rule updates, fostering scalability and handling complex, unstructured data effectively. However, it introduced challenges like interpretability needs. In GenAI evolution, this paved the way for advanced models like Transformers, impacting sectors by automating nuanced decisions. Exact extract: "The shift enabled more dynamic decision-making and adaptability with minimal manual intervention, significantly improving the efficiency of automated tasks." (Reference: Cyber Security for AI by SISA Study Guide, Section on AI Evolution and Impacts, Page 20-23).

In RAG systems, retrieving relevant information from a vector database before generation is pivotal, as it grounds responses in verified, contextually aligned data. Using embeddings and similarity metrics, the system fetches documents matching the query’s intent, ensuring accuracy and relevance. While diverse sources or feedback aid long-term improvement, the retrieval step directly drives contextual fidelity, streamlining SDLC by modularizing data access. Exact extract: "Retrieving relevant information from the vector database is crucial for ensuring contextually accurate responses in RAG systems." (Reference: Cyber Security for AI by SISA Study Guide, Section on RAG Optimization, Page 120-123).

Multi-head self-attention enhances a model's capacity to capture intricate patterns by dividing the attention process into multiple parallel 'heads,' each learning distinct aspects of the relationships within the data. This diversification enables the model to attend to various subspaces of the input simultaneously—such as syntactic, semantic, or positional features—leading to richer representations. For example, one head might focus on nearby words for local context, while another captures global dependencies, aggregating these insights through concatenation and linear transformation. This approach mitigates the limitations of single-head attention, which might overlook nuanced interactions, and promotes better generalization in complex datasets. In practice, it results in improved performance on tasks like NLP and vision, where multifaceted relationships are key. The mechanism's parallelism also aids in scalability, allowing deeper insights without proportional computational increases. Exact extract: "Multi-head attention improves learning by permitting the model to jointly attend to information from different representation subspaces at different positions, thus capturing complex relationships more effectively than a single attention head." (Reference: Cyber Security for AI by SISA Study Guide, Section on Transformer Mechanisms, Page 48-50).

Risk assessment for GenAI begins with comprehensive data flow mapping, tracing inputs, processing, and outputs to pinpoint privacy vulnerabilities like unintended data leakage. This step reveals how personal information is handled, enabling classification of risks under frameworks like GDPR or ISO 27701. It facilitates the identification of controls such as anonymization or consent mechanisms. In GenAI, where models infer from vast data, this prevents re-identification attacks. Exact extract: "A primary step in GenAI data privacy risk assessment is conducting data flow mapping to identify and mitigate privacy risks." (Reference: Cyber Security for AI by SISA Study Guide, Section on Privacy Risk Models, Page 235-238).