Fastened CROWN: Tightened Neural Network Robustness Certificates

Authors: Zhaoyang Lyu, Ching-Yun Ko, Zhifeng Kong, Ngai Wong, Dahua Lin, Luca Daniel5037-5044

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various networks trained individually verify the effectiveness of FROWN in safeguarding larger robust regions.
Researcher Affiliation Academia 1The Chinese University of Hong Kong, Hong Kong, China 2Massachusetts Institute of Technology, Cambridge, MA 02139, USA 3University of California San Diego, La Jolla, CA 92093, USA 4The University of Hong Kong, Hong Kong, China
Pseudocode No The paper describes optimization problems and procedures (e.g., projected gradient descent) but does not include a clearly labeled "Pseudocode" or "Algorithm" block with structured steps.
Open Source Code Yes Source code and the appendix is available at https://github.com/ZhaoyangLyu/FROWN.
Open Datasets Yes Experiment I. In the first experiment, we compare the improvements of FROWN and the LP-based method over CROWN on sensorless drive diagnosis networks 6 and MNIST classifiers. 6https://archive.ics.uci.edu/ml/datasets/Dataset+for+Sensorless+Drive+Diagnosis Experiment II. In our second experiment, we compute the robustness certificate on CIFAR10 networks that have 2048 neurons in each layer.
Dataset Splits No The paper mentions using specific datasets (Sensorless Drive Diagnosis, MNIST, CIFAR10) but does not provide explicit details on how these datasets were split into training, validation, or test sets (e.g., percentages, sample counts, or specific partitioning methodologies).
Hardware Specification Yes We run the LP-based method on a single Intel Xeon E5-2640 v3 (2.60GHz) CPU. We implement our proposed method FROWN using Py Torch to enable the use of an NVIDIA Ge Force GTX TITAN X GPU. However, we time FROWN on a single Intel Xeon E5-2640 v3 (2.60GHz) CPU when comparing with the LP-based method for fair comparisons.
Software Dependencies No The paper mentions using "Py Torch" for implementing FROWN and "Gurobi LP solver" but does not provide specific version numbers for these software components.
Experiment Setup Yes We leave the detailed experimental set-ups and complete experimental results to Appendix Section A.9. (From Appendix A.9): The optimization of FROWN is implemented using Adam optimizer (Kingma and Ba 2014) with a learning rate of 0.01 and 1000 iterations for each image.