EUVIMEDEUVIMED entdecken
StudienprotokollOpen Access · CC BY 4.0

MCP-HAF: PROTOCOL-DRIVEN HYBRID ANONYMIZATION FRAMEWORK FOR CONTEXT PROTECTION IN MEDICAL RAG SYSTEMS

P I Zamroz

Computer systems and network · 2026-06 · 8 (1) · 106-114

DOI: 10.23939/csn2026.01.106 ↗

Kurz gefasst

Studienprotokoll in der Zeitschrift Computer systems and network, 2026-06. Verfasst von P I Zamroz. Die Arbeit nennt 20 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid integration of Large Language Models (LLMs) into healthcare information systems poses unprecedented challenges for patient data protection, particularly within Retrieval-Augmented Generation (RAG) architectures. Traditional methods — static masking and global differential privacy — result in unacceptable semantic utility loss, critical for clinical accuracy. This paper proposes MCP-HAF (Model Context Protocol – Hybrid Anonymization Framework), a two-tier architecture implemented as middleware within the MCP protocol. Level 1 (LAM) performs on-premise semantic pseudonymization via quantized BioBERT-v1.2 NER, identifying 18 HIPAA PHI categories (F1 = 93.7%). Level 2 (AVD) applies adaptive Gaussian noise to embeddings proportional to the Local Intrinsic Dimensionality (LID) of each vector. Experiments on MIMIC-IV (40,000+ records, 4 clinical categories) demonstrate: ISR reduced 63× (88.4% → 1.4%), Recall@10 = 92.1% (+27.9 p.p. vs. Redaction, +13.6 p.p. vs. Static DP), BERTScore = 0.94, latency ≈ 57 ms. MCP-HAF satisfies all three constraints (privacy, utility, efficiency) simultaneously and ensures GDPR/HIPAA compliance. Key words: data anonymization, differential privacy, embedding inversion, GDPR/HIPAA, healthcare AI security, large language models, LID, MCP, RAG, vector perturbation.

Zitieren

APA

Zamroz, P. I. (2026). MCP-HAF: PROTOCOL-DRIVEN HYBRID ANONYMIZATION FRAMEWORK FOR CONTEXT PROTECTION IN MEDICAL RAG SYSTEMS. Computer systems and network, 8(1), 106-114. https://doi.org/10.23939/csn2026.01.106

Vancouver

Zamroz PI. MCP-HAF: PROTOCOL-DRIVEN HYBRID ANONYMIZATION FRAMEWORK FOR CONTEXT PROTECTION IN MEDICAL RAG SYSTEMS. Computer systems and network. 2026;8(1):106-114. doi:10.23939/csn2026.01.106

BibTeX
@article{Zamroz2026,
  author = {Zamroz, P I},
  title = {MCP-HAF: PROTOCOL-DRIVEN HYBRID ANONYMIZATION FRAMEWORK FOR CONTEXT PROTECTION IN MEDICAL RAG SYSTEMS},
  journal = {Computer systems and network},
  year = {2026},
  volume = {8},
  number = {1},
  pages = {106--114},
  doi = {10.23939/csn2026.01.106}
}
Kostenlos starten →