AL-AHLIYYA AMMAN UNIVERSITY. Faculty of Information Technology — Networks and Cybersecurity Department. 

Scene 1

Faculty of Information Technology — Networks and Cybersecurity Department

A Sensitivity-Aware Adaptive Differential Privacy Framework for Electronic Health Records

Presented by  Azeeza Mohammad Mustafa Aldayya

Supervisor:  Prof. Nedal Mustafa      Co-Supervisor:  Dr. Saeed Hamouda

AGENDA. Presentation Outline. 2. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 2

Sensitivity-Aware Adaptive Differential Privacy for EHRs  ·  Azeeza Aldayya

▪  Introduction & Motivation: EHRs, privacy, and the need for adaptive protection ▪  Problem Statement & Research Questions ▪  Objectives, Contributions & Research Propositions ▪  Background: Differential Privacy fundamentals & Literature Review ▪  Methodology: Sensitivity model, Composite Sensitivity Score, adaptive budgets, mechanisms ▪  Experimental Setup: dataset, preprocessing, evaluation metrics ▪  Results: CSS statistics, budget allocation, mechanism comparison ▪  Discussion, Conclusions, Limitations & Future Work

CHAPTER 1 — INTRODUCTION. Introduction & Context. 

Scene 3

▪  The rapid digitalization of healthcare has transformed how medical information is produced, stored, and used ▪  EHRs have become an essential resource for clinical decision-making, medical research, and healthcare analytics ▪  In the last two decades, healthcare establishments worldwide switched from paper records to Electronic Health Records (EHRs) ▪  Benefits realized: instant access to patient records, enhanced clinical interoperability, wide-scale medical research ▪  Outcomes: increased healthcare efficiency, fewer medical errors, and better epidemiological studies, disease surveillance, and health-policy planning

CHAPTER 1 — INTRODUCTION. Privacy & Security Concerns of EHRs. 

Scene 4

▪  EHRs hold extremely sensitive personal information: medical diagnoses, laboratory results, genetic data, mental-health data, insurance information, demographic identifiers ▪  Exposure may cause discrimination, stigmatization, monetary damage, and mistrust by the public ▪  Healthcare systems have become prime targets for cybercriminals — ransomware attacks and data breaches have grown sharply ▪  HIPAA (US) and GDPR (EU) provide governance frameworks — but legal compliance does not offer a mathematically rigorous defense       –  against modern inference attacks, dataset linkage, or re-identification methods

CHAPTER 1 — INTRODUCTION. Limitations of Existing Methods. 

Scene 5

▪  k-anonymity, l-diversity, t-closeness: useful early on, but anonymized datasets can frequently be re-identified via linkage attacks when combined with external sources — no formal privacy guarantees ▪  Fixed-budget Differential Privacy: a single global ε applied to all records and attributes — assumes all healthcare information is equally sensitive ▪  Healthcare data are heterogeneous: HIV status, mental-health conditions, or rare genetic disorders carry far more risk than a seasonal-allergy diagnosis or administrative fields ▪  Consequence: high-sensitivity data under-protected; low-sensitivity data over-perturbed, causing unjustified loss of statistical accuracy

RESEARCH GAP. Problem Statement. 6. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 6

▪  Healthcare data are complex and heterogeneous — yet most DP-based solutions use a preset, uniform global budget (ε): a one-size-fits-all approach that is impractical for real EHRs ▪  Recent adaptive DP mechanisms focus on statistical features, query patterns, or data distribution — not the sensitivity semantics of healthcare data ▪  Existing adaptive frameworks typically ignore multidimensional sensitivity factors: clinical severity, identifiability risk, data rarity, contextual healthcare importance ▪  Gap: no current solution provides an integrated framework that systematically evaluates multidimensional EHR sensitivity       –  and dynamically distributes privacy budgets based on it — while preserving a balanced privacy–utility trade-off

CHAPTER 1 — THESIS QUESTIONS. Research Questions. 

Scene 7

1.  How can a Composite Sensitivity Score (CSS) capture multidimensional privacy risks — medical, social, legal, contextual — to allocate the DP budget adaptively? 2.  How does sensitivity-aware adaptive budget allocation enhance privacy protection over traditional fixed-budget DP for records with varying sensitivities? 3.  What are the implications for key privacy metrics: re-identification risk, disclosure risk, inference-attack success rate, and privacy loss? 4.  What is the utility of the proposed framework — how does adaptive allocation affect utility compared with a fixed budget? 5.  What is the impact of adaptive ε allocation on the privacy–utility tradeoff in EHRs compared with traditional DP?

CHAPTER 1 — THESIS OBJECTIVES. Research Objectives. 

Scene 8

1.  Explore the weaknesses of current privacy-preserving methods for EHRs: conventional anonymization, fixed-budget DP, and adaptive DP approaches 2.  Build a multidimensional sensitivity assessment model with a Composite Sensitivity Score (CSS) based on Multi-Criteria Decision Analysis (MCDA) 3.  Develop the dynamic Sensitivity-Aware Adaptive Differential Privacy (SA-DP) framework with data-type-appropriate DP mechanisms while maintaining formal guarantees 4.  Evaluate via privacy metrics (re-identification risk, inference risk, privacy loss) and utility metrics (MSE, RMSE, Relative Error, correlation preservation, statistical similarity) 5.  Compare SA-DP with conventional approaches on privacy protection, data utility, and the trade-off between them

CHAPTER 1 — CONTRIBUTIONS. Thesis Contributions. 9. 

Scene 9

1.  A structured four-dimension sensitivity assessment model — medical severity, social stigma, legal implications, field-level sensitivity — quantifying sensitivity of structured EHR data 2.  An MCDA weighting approach producing a Composite Sensitivity Score (CSS): systematic, reproducible record-level sensitivity measurement 3.  An inverse mathematical relationship between CSS and the DP budget (ε), enabling dynamic allocation of privacy budgets by record sensitivity 4.  Multiple DP mechanisms — Laplace, Gaussian, Randomized Response, Hybrid — chosen by data type, preserving formal guarantees and utility 5.  Experimental comparison of adaptive vs uniform budget allocation under privacy and utility metrics 6.  A practical framework for secure EHR publication aligned with HIPAA/GDPR and privacy-by-design

CHAPTER 1 — THESIS ASSUMPTIONS. Research Propositions. 

Scene 10

1.  The SA-DP framework reduces privacy risk while preserving acceptable data utility through adaptive privacy-budget allocation 2.  Adaptive allocation provides a more balanced privacy–utility trade-off than uniform allocation under appropriate dataset conditions 3.  The Composite Sensitivity Score effectively prioritizes privacy protection according to predefined sensitivity policies 4.  The framework maintains analytical utility while implementing the selected Differential Privacy mechanisms 5.  SA-DP provides a practical and flexible privacy-preserving solution for Electronic Health Record publication

CHAPTER 2 — BACKGROUND. Background — Differential Privacy. 

Scene 11

▪  Presented in 2006 — a formal mathematical assurance: the presence or absence of one individual has a negligible impact on statistical query outputs ▪  Achieved by adding calibrated noise (typically Laplace or Gaussian) to query results ▪  Protection strength is controlled by the privacy budget ε: lower ε → stronger privacy, lower utility; higher ε → higher utility, lower privacy

Aspect Description Weakness Privacy Guarantee Provides mathematical protection against inference Depends heavily on correct ε selection Data Utility Allows statistical analysis with noise addition Accuracy decreases as privacy increases Implementation Strong theoretical foundation Technically complex and resource-intensive Repeated Queries Controlled privacy budget system Multiple queries can accumulate privacy loss

CHAPTER 2 — LITERATURE REVIEW. Literature Review & Research Gap. 

Scene 12

▪  Areas studied: ▪  Differential privacy for EHR publication; personalized / heterogeneous DP ▪  Adaptive privacy-budget allocation; composition and privacy accounting (RDP, zCDP) ▪  Semantic and regulatory sensitivity modeling; empirical privacy auditing ▪  Identified gap: ▪  No unified multidimensional sensitivity model tailored to structured EHRs ▪  Existing methods rely on single-dimension measures, manual sensitivity values, or heuristics ▪  Few approaches perform comprehensive empirical auditing of the deployed mechanisms

CHAPTER 3 — SA-DP FRAMEWORK. Methodology — Framework Pipeline. 

Scene 13

End-to-end flow of the Sensitivity-Aware Adaptive Differential Privacy (SA-DP) framework

Raw EHR Dataset 55,500 records · 26 attributes

Preprocessing cleaning · imputation · outliers

Sensitivity Model SMed · SSoc · SLeg · FSen

Composite Sensitivity Score CSS ∈ [0, 1]

Adaptive Budget ε = α / CSS  (α = 0.1)

Attribute-level Budgets numerical · temporal · categorical · financial

Noise Mechanisms Laplace · Gaussian · RR · Hybrid

Privatized Dataset published EHR with DP guarantees

Evaluation privacy metrics · utility metrics

Every record receives a Composite Sensitivity Score; the budget ε is allocated inversely to sensitivity — highly sensitive records get stronger privacy (smaller ε), less sensitive records retain more utility (larger ε)

CHAPTER 3 — EXPERIMENTAL SETUP. Dataset — Healthcare Prediction Dataset. 

Scene 14

Dataset — Healthcare Prediction Dataset

▪  Kaggle Healthcare Prediction Dataset ▪  26 original attributes · 55,500 patient records ▪  Heterogeneous data: demographics, clinical data, admissions, insurance, billing, medications, laboratory data

Attribute Name Data Type Description Name Text Patient's full name — identification purposes Age Numerical Age of the patient in years Gender Categorical Gender (Male / Female) Blood Type Categorical Blood group of the patient Medical Condition Categorical Diagnosed medical condition or illness Date of Admission Date Date the patient was admitted Doctor Text Attending physician responsible for care Hospital Text Healthcare institution providing treatment Insurance Provider Categorical Insurance company covering expenses Billing Amount Numerical Total healthcare cost incurred

CHAPTER 3 — EXPERIMENTAL SETUP. Data Preprocessing. 

Scene 15

▪  Preprocessing pipeline: ▪  Duplicate removal; missing-value imputation — mean/median for numerical, mode for categorical ▪  Outlier detection and handling — IQR and z-score methods ▪  Categorical encoding; normalization and feature scaling; feature engineering ▪  Sensitivity variables derived by the framework (not present in the original data): ▪  SMed, SSoc, SLeg, FSen — the four sensitivity dimensions ▪  Composite Sensitivity Score (CSS) and the adaptive privacy budgets (ε)

CHAPTER 3 — SENSITIVITY MODEL. Composite Sensitivity Score (CSS). 

Scene 16

▪  Four sensitivity dimensions: ▪  SMed — Medical Sensitivity: severity, chronicity, treatment complexity ▪  SSoc — Social Stigma: discrimination, reputation, psychological harm ▪  SLeg — Legal Risk: HIPAA / GDPR penalties, regulatory sanctions ▪  FSen — Field Sensitivity: attribute identifiability — IDs and genomics far more sensitive than demographics ▪  Score computation — MCDA with equal weights:

CSS  =  ω₁·SMed + ω₂·SSoc + ω₃·SLeg + ω₄·FSen        CSS ∈ [0, 1]  —  higher CSS → stronger required protection

CHAPTER 3 — SA-DP FRAMEWORK. Adaptive Privacy Budget Allocation. 

Scene 17

▪  Inverse sensitivity allocation — the core of SA-DP:

ε = α / CSS      stability constant α = 0.1 — prevents division by zero and extreme ε for low-sensitivity attributes

▪  Aggregation — record-level CSS → attribute-level budgets by data type: ▪  Numerical attributes → median CSS ▪  Temporal attributes → mean CSS ▪  Categorical attributes → maximum CSS ▪  ε clamping keeps budgets within a practical range ▪  Interpretation: small CSS → larger ε → more utility ·  large CSS → smaller ε → stronger privacy

CHAPTER 3 — SA-DP FRAMEWORK. Privacy Mechanisms & Composition. 

Scene 18

▪  Laplace mechanism (ε-DP) — numerical and continuous attributes: ages, lab values, billing costs ▪  Gaussian mechanism ((ε, δ)-DP) — temporal attributes: admission and discharge dates ▪  Randomized Response (ε-DP) — categorical attributes: blood type, diagnosis, admission type ▪  Hybrid Laplace–Gaussian — heterogeneous real-world distributions ▪  Privacy composition: ▪  Sequential composition — ε_total = Σ εⱼ over repeated queries ▪  Parallel composition — disjoint datasets split the budget safely

CHAPTER 3 — EVALUATION. Evaluation Metrics. 19. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 19

▪  Statistical fidelity: Relative Error (RE), MSE, RMSE, Wasserstein distance, Kolmogorov–Smirnov (KS) statistic ▪  Utility preservation: correlation preservation, classification accuracy ▪  Temporal coherence: mean date shift (days) ▪  Privacy assessment: re-identification risk, inference risk, privacy loss

Metric Description Interpretation Relative Error (RE) Percentage deviation between original and sanitized means Lower = better preservation Mean Squared Error (MSE) Average squared difference between original and noisy data Lower = better preservation Wasserstein Distance Difference between probability distributions Lower = higher similarity Kolmogorov–Smirnov (KS) Maximum distance between cumulative distributions Lower = stronger alignment

CHAPTER 4 — EXPERIMENTAL RESULTS. Results — Composite Sensitivity Score. 

Scene 20

Statistic Value Mean CSS 0.323 Median CSS 0.323 Standard Deviation 0.0007 Minimum CSS 0.3215 Maximum CSS 0.3231

Component Mean Score Contribution SLeg — Legal Risk 0.60 46.49% FSen — Field Sensitivity 0.442308 34.13% SMed — Medical 0.15 11.63% SSoc — Social Stigma 0.10 7.75%

▪  Very low CSS variance (σ = 0.0007) → adaptive budgets came out almost equal on this dataset ▪  SLeg dominates the score (46.49%) — the dataset is strongly regulated by billing / insurance attributes

CHAPTER 4 — EXPERIMENTAL RESULTS. Results — Adaptive Budget Allocation. 

Scene 21

Parameter Value Meaning Mean CSS 0.323077 Average sensitivity score Stability constant (α) 0.1 Prevents division by zero Adaptive privacy budget (ε) 2.363636 Budget allocated by SA-DP Formula ε = α / CSS Inverse relationship with sensitivity

Per-feature privacy parameters (Table 4-4)

Feature Category Features ε Sensitivity Δf Scale / p Mechanism Categorical Name, Gender, Blood Type, etc. 2.3636 1 0.4231 Randomized Response Numerical Age, SMed, SSoc, SLeg, etc. 2.3636 0.01 0.0042 Laplace Temporal Admission Date, Discharge Date 2.3636 1 0.4231 Gaussian Financial Billing Amount 2.3636 0.99 0.4175 Laplace

CHAPTER 4 — EXPERIMENTAL RESULTS. Results — Mechanism Comparison. 

Scene 22

Performance Metric Laplace Gaussian Hybrid Average MSE (non-financial) 2.5 2.3 2.8 Average MSE (financial) 5.2 4.8 5.5 Average KS statistic 0.2645 0.2645 0.2645 Accuracy preserved (%) 91.40% 91.40% 91.40% Mean date shift (days) 1.1 1.1 1.1 Privacy guarantee ε-DP (ε,δ)-DP Hybrid DP

▪  Randomized Response preserved > 91% of categorical information — Room Number highest at 91.57%, Doctor lowest at 91.12% ▪  Billing Amount: MSE = 0.3521 (wider numerical range) ▪  Gaussian shows the lowest numerical error (2.3); Hybrid preserves temporal coherence best

CHAPTER 4 — DISCUSSION. Discussion & Key Findings. 

Scene 23

▪  The framework implementation is correct and feasible — multidimensional sensitivity is systematically integrated into DP without compromising formal guarantees ▪  Experiment confirmed: correct computation of CSS, budget conversion, and mechanism deployment ▪  Key limitation in this experiment: CSS variability was extremely low (σ = 0.0007) — adaptive budgets ≈ equal       –  Consequently the framework behaved as a fixed-budget DP approach on this dataset ▪  The full adaptive capability must be assessed with data of higher sensitivity diversity ▪  SLeg dominated the CSS (46.49%), reflecting the dataset's regulatory/billing-sensitive attributes ▪  Laplace generally showed lower utility loss than Gaussian / Hybrid in this configuration

CHAPTER 5 — CONCLUSION. Conclusions. 24. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 24

▪  SA-DP unifies: multidimensional sensitivity assessment → MCDA-based CSS → adaptive privacy-budget allocation → formal privacy accounting → multiple DP mechanisms ▪  Experimental results confirmed the correctness and feasibility of the designed adaptive privacy framework ▪  Preserves Differential Privacy guarantees and analytical utility for structured EHR publication ▪  The adaptive advantage over fixed-budget DP could not be empirically demonstrated on a low-heterogeneity dataset — it remains to be validated on more diverse data ▪  The framework advances DP by mapping privacy strength to data sensitivity, offering a balanced, context-focused solution aligned with HIPAA/GDPR

CHAPTER 5 — LIMITATIONS. Limitations. 25. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 25

1.  Sensitivity assessment relied on synthetic data only 2.  Simplified CSS: equal weights and assumed independence among the sensitivity dimensions — unlikely to reflect real-world situations 3.  No privacy testing against adversaries — membership or attribute-inference attacks — so real-world privacy guarantees remain unverified 4.  Evaluation done on relatively small datasets; not fitted for real-time and large-scale streaming scenarios 5.  Low CSS variability in the chosen dataset limited the demonstration of adaptivity — budgets were nearly uniform

CHAPTER 5 — FUTURE WORK. Future Work. 26. Sensitivity-Aware Adaptive Differential Privacy for EHRs · Azeeza Aldayya. 

Scene 26

1.  Practical assessment with large real-world EHR datasets and better synthetic data with improved sensitivity diversity 2.  Extend SA-DP to federated learning, deep learning (DP-SGD), and other data formats — textual, image, and time-series — in healthcare 3.  Develop a dynamic privacy-budget allocation system with explainability and human-in-the-loop oversight

Thank You. Questions & Discussion Welcome. Azeeza Mohammad Mustafa Aldayya. 

Scene 27

Supervisor: Prof. Nedal Mustafa   ·   Co-Supervisor: Dr. Saeed Hamouda

Al-Ahliyya Amman University — Networks and Cybersecurity Department  ·  2026

SA-DP_Thesis_Defense_NavyGold

Scene 1 (0s)

Scene 2 (16s)

Scene 3 (38s)

Scene 4 (1m 4s)

Scene 5 (1m 28s)

Scene 6 (1m 55s)

Scene 7 (2m 23s)

Scene 8 (2m 53s)

Scene 9 (3m 24s)

Scene 10 (3m 56s)

Scene 11 (4m 20s)

Scene 12 (4m 48s)

Scene 13 (5m 9s)

Scene 14 (5m 43s)

Scene 15 (6m 12s)

Scene 16 (6m 33s)

Scene 17 (6m 55s)

Scene 18 (7m 18s)

Scene 19 (7m 38s)

Scene 20 (8m 4s)

Scene 21 (8m 35s)

Scene 22 (9m 8s)

Scene 23 (9m 39s)

Scene 24 (10m 7s)

Scene 25 (10m 33s)

Scene 26 (10m 57s)

Scene 27 (11m 16s)