Comparative_AI_in_CSComparative_AI_in_CS.pptx

Comparative_AI_in_CS

1. Comparative Study of AI in Cyber Security & Healthcare Summarized Research Papers (2023–2024)

2. AI, Machine Learning and Deep Learning in Cyber Risk Management Year: 2023 Problem: Static/manual cyber risk management struggles against evolving threats in healthcare and critical infrastructures. Methods: Survey of AI/ML/DL (SVM, neural networks, Bayesian, hybrid models). Results: DL and hybrid models improve predictive risk analysis and anomaly detection over classical ML. Merits: Comprehensive mapping; automation and adaptability emphasized. Demerits: Conceptual; limited real-world validation. Scope: Lightweight, interpretable AI-driven risk management tailored to EHRs/IoMT with clinical validation.

3. Anomaly Detection Model in Network Security Situational Awareness Year: 2024 Problem: Traditional IDS fails to adapt to evolving/zero‑day threats in healthcare networks. Methods: Survey of supervised, unsupervised, and deep learning (CNN/RNN, autoencoders). Results: DL generally outperforms classic ML; unsupervised helps with zero‑day; hybrids improve adaptability. Merits: Structured taxonomy; healthcare‑relevant insights. Demerits: No implementation; survey‑only. Scope: Federated + explainable deep anomaly detection for healthcare IoT/wearables.

4. Cryptographic Primitives in Privacy‑Preserving Machine Learning: A Survey Year: 2024 Problem: Balancing patient data privacy with ML utility in sensitive healthcare settings. Methods: Survey of HE, MPC, DP, ZK‑proofs for ML pipelines. Results: HE strongest privacy but costly; DP efficient with accuracy loss; MPC balanced but complex. Merits: Clear taxonomy of PPML trade‑offs. Demerits: No experimental/healthcare deployment. Scope: Lightweight PPML for medical IoT; DP‑aware federated learning for multi‑hospital EHR.

5. Deep Reinforcement Learning for Cyber Security Year: 2023 Problem: Static defenses can’t keep up with adaptive attackers in healthcare/IoT networks. Methods: Survey + case studies of DRL agents for IDS/IPS and automated response. Results: Higher detection/adaptability vs classic ML; effective attack‑defense dynamics. Merits: Dynamic, policy‑learning approach; promising adaptability. Demerits: Data/computation hungry; deployment complexity. Scope: Lightweight/edge DRL for IoMT; explainable DRL with safety constraints for clinics.

6. Detection of DoS Attack in Wireless Sensor Networks: A Lightweight ML Approach Year: 2023 Problem: Resource‑constrained healthcare WSNs vulnerable to DoS; need efficient detection. Methods: Decision Trees, Logistic Regression on WSN datasets. Results: >95% detection accuracy with low memory/CPU use. Merits: Lightweight; suitable for sensors/wearables. Demerits: Covers only DoS; limited datasets. Scope: Extend to multi‑attack detection; real‑world trials on medical wearables.

7. Ensemble Adaptive Online ML in Data Stream: Case Study in Intrusion Detection Year: 2024 Problem: Batch IDS struggles with continuous healthcare data streams (IoMT/hospital networks). Methods: EnsAdp_CIDS—ensemble adaptive online learning on CICIDS‑2017, CIC‑IoT‑2023, CIC‑MalMem‑2022. Results: Accuracies: 99.77%, 98.93%, 99.85%; outperforms baselines. Merits: High accuracy; adapts in real‑time. Demerits: Computationally intensive for edge devices. Scope: Edge/fog‑optimized adaptive IDS for hospitals; explainability for compliance.

8. IP2FL: Interpretation‑Based Privacy‑Preserving Federated Learning for ICPS Year: 2024 Problem: Need privacy, efficiency, and interpretability in FL for sensitive data (applicable to healthcare). Methods: Additive HE + dual feature selection + Shapley‑value explanations in FL. Results: Maintains accuracy with reduced overhead; interpretable client contributions. Merits: Balances privacy, efficiency, interpretability. Demerits: Complex to deploy; scalability not fully tested. Scope: Federated EHR across hospitals with AHE + XAI; optimize comms for low‑bandwidth clinics.

9. ML‑Based Cyber Threat Detection with Explainable AI Insights Year: 2024 Problem: Lack of transparency in ML models for malware/ransomware detection in healthcare. Methods: SVM/DT/KNN/RF with SHAP & LIME on malware dataset. Results: RF achieved ~100% on dataset; XAI clarifies feature importance. Merits: High accuracy + interpretability. Demerits: Single dataset; overfitting risk. Scope: XAI‑first ransomware detection on real hospital telemetry; robust evaluation.