Adversarial Exposure Validation (AEV) is a technique used mainly in machine learning and data science to detect dataset shift or data leakage between different datasets (usually training vs. validation/test datasets). It helps determine whether the model might perform unrealistically well because the datasets are not distributed the same way.

Core Idea

AEV trains a secondary model whose job is to answer one question:

“Can a model tell whether a sample comes from dataset A or dataset B?”

For example:

• Dataset A → Training data

• Dataset B → Validation/Test data

You label them like this:

Then you train a classifier to distinguish them.

How It Works

1. Combine datasets

   • Merge training and validation/test data.

2. Add source labels

   • Training samples → label 0

   • Validation/test samples → label 1

3. Train a classifier

   • Any classifier works (e.g., logistic regression, random forest, gradient boosting).

4. Measure performance

   • Evaluate using metrics like AUC.