Model Theft
Unauthorized extraction or copying of the model.
What is Model Theft?
Model Theft involves unauthorized access to, copying of, or extraction of an LLM's weights, architecture, or functionality.
Attackers can steal models by:
- Direct Access: Gaining physical or network access to the server hosting the model.
- Extraction Attacks: Querying the model repeatedly to reconstruct its parameters or train a shadow model that mimics its behavior.
- Side-Channel Attacks: Analyzing power consumption or timing data to infer model details.