Please enable JavaScript.
Coggle requires JavaScript to display documents.
Distribution Invariant Differential Privacy - Introduction - Coggle Diagram
Distribution Invariant Differential Privacy - Introduction
Data Privacy In Big Data Era
Increasing importance in many fields
Massive amount of sensitive information
Digitally collected, stored, transferred, and analyzed
Adoption of Differential Privacy
Biomedical Research (References: [31, 33, 34, 37])
Technology Companies
Google (Reference: [22])
Apple (Reference: [3])
Microsoft (Reference: [13])
LinkedIn (Reference: [41])
Amazon (Reference: [11])
Facebook (Reference: [47])
Sociology (Reference: [50])
Epidemiology (Reference: [57])
U.S. Decennial Census (2020)
First-time use of differential privacy
Protects confidentiality of individuals (Reference: [53])
Differential Privacy
Improves transparency and accessibility in AI (Reference: [28])
Useful for publicly released data (e.g., census and survey data)
Protects extreme observations
Quantifies privacy (Reference: [17])
Differential Privacy Methods
Developed by scientists
Goal: Protect data privacy and promote data sharing
Benefits of Data Sharing
Integration of privatized data from multiple entities
Strengthens data analysis
Economic and societal benefits in the big data era (Reference: [56])
Trade-off between Privacy and Accuracy
Data privatization can alter analysis results
Trade-off between statistical accuracy and differential privacy
Sacrificing accuracy for privacy (References: [5, 26, 27, 50])
Reconciliation of Privacy and Accuracy
Preservation of the original data's distribution
Distribution-Invariant Privatization Method (DIP)
Transform and perturb data
Suitable transformation to recover the original distribution
Maintains statistical accuracy
Achieves differential privacy at desired levels of protection
Differential Privacy Literature
Computer Science
Research and methods
Statistics
Statistical implications and approaches
Major Directions in Differential Privacy
Computer Science
Achieves privacy protection through mechanisms
Examples
Laplace mechanism (Reference: [19, 20])
Exponential mechanism (Reference: [43])
Minimax optimal procedures (Reference: [15])
Statistics
Achieves differential privacy for models/algorithms
Examples
Deep learning (Reference: [1])
Boosting (Reference: [21])
Stochastic gradient descent (Reference: [2])
Risk minimization (Reference: [7])
Random graphs (Reference: [38])
Function estimation (Reference: [30])
Parametric estimation (Reference: [4])
Regression diagnostics (Reference: [8])
Top-k selection (Reference: [16])
General Statistical Framework for Differential Privacy
Introduced (Reference: [58])
Investigates theoretical properties of privatization mechanisms (References: [36, 56])
Types of Differential Privacy
Relaxed or approximate differential privacy (References: [1, 2, 6, 18])
Local differential privacy (References: [13, 23, 39, 48])
Random differential privacy (Reference: [29])
Renyi differential privacy (Reference: [44])
Gaussian differential privacy (Reference: [14])
Challenges in Data Privacy
Existing mechanisms alter data distribution
Negative values in count data
Data-dependent post-processing required
Distribution-Invariant Privatization (DIP) Mechanism
Addressing challenges for all types of data
Satisfies differential privacy
Approximately preserves the original distribution
Scalable to massive/high-dimensional data
Benefits of DIP
Maintains statistical accuracy
No trade-off between accuracy and privacy
Enables various data analysis tasks without sacrificing accuracy
Regression
Classification
Graphical models
Clustering
Other statistical and machine learning tasks