Co(ve)rtex: Machine Learning Models as Storage Channels and Their (Mis-)Applications

We introduce Co(ve)rtex, a framework demonstrating that modern over-parameterized machine learning models can function as covert storage channels, capable of embedding and retrieving large amounts of hidden information without impacting primary task accuracy.

Our results reveal a novel security and privacy risk in deployed ML systems, enabling stealthy data exfiltration and misuse. We further propose error-correction–based defenses to mitigate these threats, highlighting the need for new safeguards when deploying large-capacity models in security- and privacy-sensitive settings.