Researchers discuss the possibility of attackers extracting sensitive information from machine learning models without knowing the model's parameters, and share their findings on a defense method with positive results.