Model theft
Model theft is when someone steals a company's private AI 'brain'—either by copying the secret files directly or by asking the AI so many questions that they can build a cheap, fake version of it.
Model theft involves the unauthorized acquisition of a proprietary model's intellectual property, achieved either by exfiltrating model weights from a secure environment or by using query-based distillation to train a surrogate model that mimics the original's behavior.
Model theft is the illicit extraction or replication of a proprietary machine learning model's parameters or functional mapping. This is executed via direct exfiltration of model artifacts or through adversarial distillation techniques, resulting in the misappropriation of intellectual property and the creation of a surrogate model that facilitates downstream adversarial exploitation.
evolution
- 2016-03 · historyModel Extraction Attacks
Researchers demonstrated that machine learning models could be replicated by querying them to train a substitute model, effectively stealing their functionality.
- 2020-09 · historyStealing Hyperparameters
Studies showed that attackers could infer sensitive model architecture and hyperparameters through side-channel analysis of query responses.
- 2022-12 · historyLLM Distillation Attacks
The rise of large language models led to widespread concerns regarding 'model stealing' via API distillation, where smaller models are trained to mimic proprietary LLM outputs.
- 2023-05 · historyWeight Exfiltration Risks
The leak of Meta's LLaMA weights highlighted the shift from query-based distillation to direct exfiltration of model parameters as a primary security threat.