Training Provably Robust Models by Polyhedral Envelope RegularizationDownload PDF

25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • Keywords: deep learning, adversarial attack, robust certification
  • Abstract: Training certifiable neural networks enables one to obtain models with robustness guarantees against adversarial attacks. In this work, we use a linear approximation to bound model’s output given an input adversarial budget. This allows us to bound the adversary-free region in the data neighborhood by a polyhedral envelope and yields finer-grained certified robustness than existing methods. We further exploit this certifier to introduce a framework called polyhedral envelope regular- ization (PER), which encourages larger polyhedral envelopes and thus improves the provable robustness of the models. We demonstrate the flexibility and effectiveness of our framework on standard benchmarks; it applies to networks with general activation functions and obtains comparable or better robustness guarantees than state-of-the-art methods, with very little cost in clean accuracy, i.e., without over-regularizing the model.
10 Replies