AnomiGAN

AnomiGAN anonymizes personal medical data using Generative Adversarial Networks to preserve privacy while retaining predictive utility for research.

Key Features:

• Privacy Preservation: Provides privacy protection comparable to differential privacy (DP) by transforming personal medical data into anonymized representations.
• High Prediction Performance: Maintains high prediction performance of downstream models despite anonymization compared with many traditional techniques.
• Trade-off Management: Exposes a controllable trade-off between privacy preservation and prediction accuracy and provides mechanisms to adjust this balance for different applications.
• Mathematical Foundation: Supplies a rigorous mathematical overview that underpins the anonymization model.
• Validation and Utility: Validated using datasets from the UCI Machine Learning Repository to demonstrate effectiveness on real-world data.
• Implementation Libraries: Implemented using Scikit-learn (version 0.18) and Keras (version 2.0.6).

Scientific Applications:

• Genomic Data Protection: Anonymizes short DNA sequences that could potentially identify individuals or their relatives to enable safer sharing and storage of genomic data.
• Disease Prevention and Treatment Research: Enables analysis of medical datasets for research in disease prevention and treatment while preserving individual privacy.

Methodology:

AnomiGAN uses a generative adversarial network architecture in which a generator creates anonymized versions of input data and a discriminator evaluates them against real data, with the adversarial training loop iteratively improving anonymization while retaining essential predictive patterns.