Course contentsShow
← Machine Learning and Deep Learning
Lesson 3331 of 3,53873. Privacy in Machine LearningFree lesson

Attribute Inference Attacks

Study attacks that infer sensitive attributes about training data individuals not directly used as model features.

Attribute Inference Attacks

What you'll learn: How adversaries can infer sensitive personal attributes about individuals in training data, even when those attributes weren't directly used as model features.

The Core Idea

Attribute inference attacks exploit correlations in data to reveal sensitive information about training set individuals that the model never explicitly learned. Unlike membership inference (which asks "was this person in the training data?") or model inversion (which reconstructs training samples), attribute inference asks: "what is this person's sensitive attribute X?"

Think of it like a detective puzzle: if a hospital uses patient data to train a disease prediction model but excludes HIV status as a feature, an attacker might still infer HIV status by observing how the model responds to other correlated attributes (age, zip code, specific medication patterns). The model inadvertently "memorizes" these correlations during training.

How It Works

An adversary with query access to your model can:

  1. Identify a target individual known to be in the training set
  2. Query the model with variations of that person's known attributes
  3. Use the model's predictions to infer their unknown sensitive attribute (political affiliation, health condition, income bracket)

The attack succeeds because models learn joint distributions—even "irrelevant" features can leak information about excluded sensitive attributes through proxy variables and hidden correlations in the training data.

Why This Matters

This is particularly dangerous because organizations often believe removing a sensitive attribute from features provides privacy. But correlation structures remain embedded in the model's learned representation, making inference attacks possible even against seemingly "privacy-aware" systems.

Key Takeaway: Attribute inference attacks demonstrate that simply excluding sensitive attributes from model inputs doesn't guarantee privacy—attackers can exploit learned correlations to infer protected information about training data subjects.