Reproduce “Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jun Guo, Aishan Liu, Xingyu Zheng, Siyuan Liang, Yisong Xiao, Yichao Wu, Xianglong Liu

Researchers at Beihang University developed InI (Isolation and Induction), a training framework that enhances deep neural network robustness against model stealing attacks by embedding defenses directly into model parameters. The method significantly reduces inference overheads, achieving up to 25.4 times faster inference than prior defenses, while maintaining high benign accuracy and reducing clone model accuracy, for example, by substantial margins on datasets like CIFAR-100.

Key insights

  • Deep neural networks can be trained to acquire defensive capabilities against model stealing as part of their inherent functionality, leveraging over-parameterization to embed robustness.
  • Differentiating between in-distribution benign queries and out-of-distribution malicious queries allows models to selectively provide accurate outputs to legitimate users while misleading adversaries.
  • By disrupting the adversary's gradient-based learning process and minimizing information gain, a victim model can significantly reduce the effectiveness of model stealing attacks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.