Towards Fast Computation of Certified Robustness for ReLU Networks

Authors: Lily Weng, Huan Zhang, Hongge Chen, Zhao Song, Cho-Jui Hsieh, Luca Daniel, Duane Boning, Inderjit Dhillon

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that (1) our methods deliver bounds close to (the gap is 2-3X) exact minimum distortions found by Reluplex in small networks while our algorithms are more than 10,000 times faster; (2) our methods deliver similar quality of bounds (the gap is within 35% and usually around 10%; sometimes our bounds are even better) for larger networks compared to the methods based on solving linear programming problems but our algorithms are 3314,000 times faster; (3) our method is capable of solving large MNIST and CIFAR networks up to 7 layers with more than 10,000 neurons within tens of seconds on a single CPU core. In this section, we perform extensive experiments to evaluate the performance of our proposed two lower-bound based robustness certificates on networks with different sizes and with different defending techniques during training process.
Researcher Affiliation Academia 1Massachusetts Institute of Technology, Cambridge, MA 2UC Davis, Davis, CA 3Harvard University, Cambridge, MA 4UT Austin, Austin, TX.
Pseudocode Yes We list our complete algorithm, Fast-Lin, in Appendix D. We list our full procedure, Fast-Lip, in Appendix D.
Open Source Code Yes https://github.com/huanzhang12/Certified Re LURobustness
Open Datasets Yes our method is capable of solving large MNIST and CIFAR networks up to 7 layers with more than 10,000 neurons within tens of seconds on a single CPU core.
Dataset Splits No The paper mentions using "100 random test images" but does not specify training, validation, or test split percentages or detailed methodology for data partitioning.
Hardware Specification No running time (per image) for all methods is measured on a single CPU core. While 'single CPU core' is a hardware detail, it lacks specific model numbers or types.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or solver versions) needed to reproduce the experiments.
Experiment Setup Yes Table 2. Comparison of the lower bounds for distortion found by our algorithms on models with defensive distillation (DD) (Papernot et al., 2016) with temperature = 100 and adversarial training (Madry et al., 2018) with = 0.3 for three targeted attack classes.