Preventing Model Extraction in Contextual Linear Optimization

Key points

• Our focus is on how to prevent attackers, who can submit context queries, from extracting the underlying prediction model.1
• For linear models, we provide sufficient conditions where the attacker can learn the model up to positive scaling by only observing the decision.1
• These conditions motivate an adaptive attack that alternates inverse optimization with boundary sampling, i.e., using context queries leading to decision boundaries of the current attacker model.1
• To prevent extraction, we propose a mechanism that perturbs decisions only when the predicted cost vector is near a decision boundary.1

Sources (1)

• [1]Preventing Model Extraction in Contextual Linear Optimization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:55 PM

  Consider a contextual linear optimization problem, where a predict-then-optimize (PTO) system returns optimal decisions without revealing the cost vectors used in its downstream linear program.

  Our focus is on how to prevent attackers, who can submit context queries, from extracting the underlying prediction model.

Extractive summary: sentences quoted from the sources.