Membership Inference Attack

Understanding membership inference attacks - a privacy attack that determines whether a specific data point was used in training a machine learning model.

What is a Membership Inference Attack?

A membership inference attack (MIA) is a privacy attack where an adversary attempts to determine whether a particular data record was part of the training dataset used to build a machine learning model.

How It Works

  1. Shadow Model Training - Train multiple models on similar data distributions
  2. Attack Model Training - Train a classifier to distinguish member vs non-member behavior
  3. Inference - Use the attack model to classify target records

Why It Matters

  • Privacy regulations (GDPR, HIPAA)
  • Sensitive data exposure
  • Model vulnerability assessment

Defenses

  • Differential privacy training
  • Regularization techniques
  • Output perturbation
  • Membership inference game training

Further Reading

Coming soon.