AI for Cybersecurity_ Research and Practice
Unknown
Questions & Answers from this book
Questions and answers are connected to the referenced book and its available source material.
Chapter 2: LLMs on Support of Privacy and Security of Mobile Apps: State-of-the-art and Research Directions
Chapter 3: Machine Learning-based Intrusion Detection Systems: Capabilities, Methodologies, and Open Research Challenges
What are the main open research challenges in intrusion detection for encrypted traffic as discussed in the chapter?
The chapter lists five open research challenges: optimizing detection performance, especially the high false positive rate caused by lack of content visibility; feature engineering for encrypted traffic given limited payload visibility; handling data imbalance because encrypted malicious traffic is rare and hard to collect; developing unsupervised detection for novel threats; and adapting to rapidly evolving encryption protocols and attack methods.
What are the main open research challenges for XAI in intrusion detection systems?
The main open research challenges for XAI in intrusion detection systems are addressing instability and uncertainty in explanations, adapting to dynamic network environments, moving from local to global explanations, ensuring robustness against adversarial attacks, defining tailored explainability standards for IDSs, and developing user-centric explanation frameworks for different stakeholders.
Chapter 5: Enhancing Threat Detection and Response with Generative AI and Blockchain
What are the limitations of text-only RAG systems, and how do graph-based RAG systems address these limitations?
Text-only RAG systems retrieve based heavily on textual similarity, so they struggle to capture latent or complex relationships that require multi-hop reasoning across datasets or external sources like ontologies and knowledge graphs. This leads to lower retrieval accuracy and can pull in irrelevant documents, limiting the generative model's effectiveness. Graph-based RAG systems address these limitations by representing the knowledge base as a graph of entities and relationships, letting the retriever traverse connections and uncover relationships between disparate pieces of information for richer and more relevant retrieval.
How does the RAG-based approach using a knowledge graph generate responses to user queries?
When a user asks a question, the RAG-based system first converts the question into a Cypher query, Neo4j's native query language. It then executes that query against the Neo4j knowledge graph to retrieve relevant nodes and relationships for cybersecurity entities such as CVE, CWE, CAPEC, and ATT&CK. The query output is preprocessed so the most relevant structured data becomes context, and that context is fed into a large language model to synthesize a final response.
Chapter 8: Machine Learning Attacks on Signal Characteristics in Wireless Networks
According to Table 6.4, which privacy-preserving technologies provide full protection for local models and partial protection for global models?
Homomorphic encryption and trusted execution environments (TEEs) provide full protection for local models and partial protection for global models, according to Table 6.4.
Why are clients more likely than servers to act as model downgrade attackers?
Clients are more likely to act as model downgrade attackers because they can benefit from sabotaging the training process, such as undermining the utility of the shared model for competing clients. Servers are less likely to do so because they typically gain no advantage from such actions.
Chapter 9: Secure by Design
List and describe three architectural tactics from the 'Withstand Cyber Events' category that can be used to proactively resist attacks.
Three Withstand Cyber Events tactics are validate inputs, authenticate actors, and limit exposure. Validate inputs sanitizes, neutralizes, and validates any externally provided inputs to prevent malformed data or code injection. Authenticate actors verifies that an actor is indeed who it claims to be. Limit exposure minimizes the attack surface by designing the system with the least needed number of entry points.
What are the limitations of Byzantine-resilient defenses against adaptive model-poisoning attacks?
Byzantine-resilient defenses generally assume that attackers are naive and form a minority, so adaptive model-poisoning attackers can undermine them by adjusting updates to evade detection. Even a small fraction of collaborating adaptive attackers can bypass Byzantine-robust aggregation and anomaly detection, especially through backdoor attacks that preserve overall accuracy. Stronger countermeasures often require impractical assumptions such as server access to a root dataset or client update histories, creating privacy and efficiency trade-offs.
Chapter 10: DDoS Detection in IoT Environments: Deep Packet Inspection and Real-world Applications
Chapter 12: AI Implications for Cybersecurity Education and Future Explorations
According to CSEC2017, what is the purpose of the Cybersecurity Essentials within each Knowledge Area?
According to CSEC2017, the Cybersecurity Essentials within each Knowledge Area capture the cybersecurity proficiency that every student needs to achieve regardless of program focus. They should be introduced early and reinforced throughout every cybersecurity program.
List the student learning outcomes for the Safe and Secure AI area.
The student learning outcomes for Safe and Secure AI are: 1. Understand the legal and ethical implications of AI security and privacy. 2. Understand AI and machine learning fundamentals. 3. Assess AI system vulnerabilities. 4. Secure AI pipelines. 5. Defend against AI-driven adversarial attacks. 6. Ensure data integrity in AI applications. 7. Mitigate bias and fairness issues in AI systems. 8. Design robust AI systems. 9. Monitor AI system performance.
Chapter 15: AI for Android Malware Detection and Classification
How does the attention mechanism in the MLP-SVM framework enhance feature extraction for Android malware detection?
The attention mechanism dynamically assigns weights to features based on their importance for malware detection, causing the MLP to focus on the most discriminative attributes. This refines the feature representations, improving the model's ability to distinguish benign from malicious applications and enhancing interpretability. The attention-weighted features are then passed to LDA for dimensionality reduction while preserving informative content.
According to the SHAP analysis in Section 15.5.3, which components have the highest mean absolute SHAP values, and what does the asymmetric distribution of SHAP values indicate about malware versus benign classification?
Components 1 and 2 have the highest mean absolute SHAP values, exceeding 1.0, and Component 14 is the third most influential; Components 8-12 are relatively minor. The asymmetric distribution, with long negative tails, indicates that certain feature combinations strongly point to benign applications, while malware detection relies on more subtle combinations of positive indicators.
What are the three loss functions used to train the reconstruction module in the proposed method, and what is the purpose of each?
The three loss functions are decision-level loss, feature-level loss, and L2 loss. Decision-level loss combines score loss and drift loss to align the tracker's classification and regression outputs; feature-level loss reduces the distance between feature maps of reconstructed and original frames; L2 loss minimizes pixel-wise difference between reconstructed and original frames.
Chapter 17: AI-powered Physical Layer Security in Industrial Wireless Networks
Chapter 18: The Security of Reinforcement Learning Systems in Electric Grid Domain
How is the reward function defined for the DQN agent in the microgrid secondary control case study, and what is the agent's objective?
In the microgrid case study, the DQN agent's reward is r = -(delta-omega) - (delta-v), where delta-omega is the sum over the three DGs of (omega_ss - omega_i) and delta-v is the sum over the three DGs of (v_ss - v_i). The agent's objective is to maximize this reward, which equivalently minimizes each DG's frequency and voltage deviation from the steady-state reference values under stochastic conditions.
How does the proposed PLS-enhanced ZT security framework integrate physical layer security techniques to enhance Zero Trust in wireless IIoT?
The PLS-enhanced Zero Trust (PLS-ZT) framework for wireless IIoT applies a three-step strategy that maps physical layer security techniques onto Zero Trust functions: security zone formation, device authentication, and cryptographic negotiation. It uses artificial noise and beamforming to form isolated security zones and secure tunnels, RF and channel fingerprinting for mutual authentication, and physical layer key distribution for lightweight cryptographic negotiation. These techniques together implement the Zero Trust requirements of verifying every device before trust, protecting legacy devices, defining flexible zone boundaries, and securing data flows between OT and IT systems.
Chapter 19: Geopolitical Dimensions of AI in Cybersecurity: The Emerging Battleground
Chapter 20: Robust AI Techniques to Support High-consequence Applications in the Cyber Age
Why does the chapter caution against referring to the output pi of a neural network trained with BCE loss as a probability?
The chapter cautions against calling pi a probability because a BCE-trained network minimizes the discrepancy between its output and the binary label, not the true probability distribution of that label. Treating pi as a true probability requires idealized assumptions such as infinite data and perfect calibration, which rarely hold in practice. Instead, pi is best described as a classification decision or confidence-like score, not an exact probability.
How does the chapter propose to modify empirical risk minimization (ERM) to incorporate SME knowledge?
The chapter proposes encoding SME knowledge as a vector v(SME) in the same feature space as the data or its embedded form, then modifying ERM so the model is trained to respect that knowledge. This is operationalized through a directional derivative D(SME)_k = ∇f_k(g(x)) · v(SME), which provides a gradient that optimization algorithms can use during training.
What are the main causes of performance hazards in foundation models, and what techniques can be used to mitigate them?
Performance hazards in foundation models are mainly caused by biases in the training data, lack of diversity in the data, or insufficient training of the model. They can be mitigated with data augmentation, debiasing techniques, active learning, adversarial training, and regular evaluation and updating of the model's architecture, algorithms, and training data.
You may also be interested in
FastAPI: Modern Python Web Development
Bill Lubanovic;
37 questions
Generative Deep Learning: Teaching Machines to Paint, Write, Compose, and Play
David Foster;
33 questions
AI for Time Series_ Volume 1_ Unlocking Patterns with Deep Learning
Min Wu;Emadeldeen Eldele;Zhenghua Chen;Shirui Pan;Qingsong Wen;Xiaoli Li;
25 questions
AI ChatBots For Dummies
Eric Butow, Kelly Noble Mirabella
17 questions