Train Proxy via Replication
This is when someone copies a private AI by asking it thousands of questions and recording the answers. They then use those answers to teach their own 'copycat' AI to act just like the original, allowing them to study it in private to find ways to break or trick it.
An adversary replicates a target model by querying its inference API to build a labeled dataset of inputs and outputs. This dataset is then used to train a surrogate model that approximates the target's decision boundaries, enabling the adversary to perform offline analysis and craft adversarial examples without further interaction with the victim system.
Train Proxy via Replication is a model extraction technique where an adversary systematically queries a victim's inference API to construct a synthetic training set. By distilling the target model's functional behavior into a local surrogate model, the adversary gains a white-box or gray-box environment to optimize adversarial perturbations, facilitating downstream attacks such as evasion or poisoning while minimizing detection risk and API costs.