Assessing and improving the robustness of Binarized Neural Networks
(2023) Workshop VeriLearn — Location: Krakow, Poland (30.September.2023)
Files
robust___BNNs.pdf
Open Access - Adobe PDF
- 1.4 MB
Details
- Authors
- Abstract
- Several recent publications have studied the robustness of classical neural networks and how to improve robustness thanks to adversarial training. These techniques are however hard to apply on Binarized Neural Networks (BNNs). In this work, we address this challenge and introduce a method to perform adversarial training on BNNs, by using a recent SAT-solver based technique for finding adversarial examples. This method is able to increase the robustness of BNNs according to a number of different metrics: adversarial accuracy, time taken by the considered attack and resistance of the model to this attack. It can also be generalized easily to other methods of verification and other types of models as it is a simple extension of the classical training algorithm.
- Affiliations
Citations
Ronval, B. (2023). Assessing and improving the robustness of Binarized Neural Networks. Workshop VeriLearn, Krakow, Poland. https://hdl.handle.net/2078.5/269897