A Tool to Detect Privacy Leaks in Deep Learning Models

Context & Background

Deep Learning (DL) models are trained on massive public and private datasets. However, research shows that these models act like black boxes that memorize training data. This makes them vulnerable to privacy attacks where adversaries reconstruct raw training records or identify patient records used in clinical models.

Problems to be Addressed

Adversaries can launch membership inference attacks, model inversion, and model stealing. If sensitive health or financial data is exposed, organizations face legal lawsuits and trust loss. Existing defense methods degrade model accuracy.

Aims and Objectives

1. Investigate privacy leaks in black-box and white-box models.
2. Formulate benchmarking metrics to score model vulnerabilities.
3. Design a diagnostic tool that scans models for leaks before deployment.

Methodology

The team develops attack simulations (membership inference, feature perturbation) across multiple DL architectures. Causal inference and statistical correlation tests analyze how model outputs vary with input updates. Differential privacy methods are integrated to build robust defense benchmarks.

Expected Outcomes

A diagnostic software tool scoring ML model privacy, APIs for privacy-preserving training, and publications in leading machine learning venues.