How API Outputs Change Extraction Risk: Labels vs. Probabilities
Explore how different API response formats—labels, probabilities, and richer outputs like embeddings—influence the amount of information revealed about a machine learning model. Understand how these outputs affect the ease of model extraction and the associated security risks. This lesson helps you evaluate and design API responses to balance functionality and confidentiality.
You run the same image-classification prediction API three ways, and the only thing that changes is what the response contains. For the same input image, one version returns a single class label, another returns a probability or confidence vector, and a third returns something richer like logits, embeddings, or explanations.
If the goal is to functionally reproduce the model behind the API, which response format gives the most leverage per query, and what would count as evidence for that choice? The important ...