Model Inversion / Reconstruction Attack

An attack that abuses access to a trained decoder to reconstruct representative inputs or recover a private attribute, effectively running the model backward. Against a neural decoder this could mean regenerating characteristic neural patterns associated with a class, or inferring a sensitive label (a condition, an intended word) from the model together with limited outputs.

Reconstruction is generally approximate rather than a verbatim recovery of raw signals, and its success depends on how much the model has entangled private information with the task it was trained for. It is a central reason not to treat a released decoder or its gradients as safe to share even when the raw recordings are withheld.