Decision Boundary Analysis of Adversarial Examples

Authors: Warren He, Bo Li, Dawn Song

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

Reproducibility Variable Result LLM Response
Research Type Experimental Additionally, we report similar experimenal results on a small subset of Image Net in Appendix D. We use two popular academic image classification datasets for our experiments: MNIST, consisting of black-and-white handwritten digits (Le Cun, 1998), and CIFAR-10, consisting of small color pictures (Krizhevsky & Hinton, 2009).
Researcher Affiliation Academia Warren He, Bo Li & Dawn Song Computer Science Division University of California, Berkeley
Pseudocode No The paper describes the algorithms and network architecture in text and diagrams (Figure 4) but does not include formal pseudocode blocks.
Open Source Code Yes We have released the code we used at https://github.com/sunblaze-ucb/ decision-boundaries.
Open Datasets Yes We use two popular academic image classification datasets for our experiments: MNIST, consisting of black-and-white handwritten digits (Le Cun, 1998), and CIFAR-10, consisting of small color pictures (Krizhevsky & Hinton, 2009). Additionally, we report similar experimenal results on a small subset of Image Net (Russakovsky et al., 2015) in Appendix D.
Dataset Splits No For MNIST, we train on 8,000 examples (each example here contains both a benign image and an adversarial image) for 32 epochs, and we test on 2,000 other examples. For CIFAR-10, we train on 350 examples for 1,462 epochs, and we test on 100 other examples.
Hardware Specification Yes In our slowest attack, on the PGD adversarially trained CIFAR-10 model, our attack takes around 8 minutes per image on a Ge Force GTX 1080. In our slowest experiment, with benign images on the PGD adversarially trained wide Res Net w32-10 CIFAR-10 model, it took around 70 seconds per image to compute decision boundary information for 1,000 directions on a Ge Force GTX 1080.
Software Dependencies No The paper mentions specific methods like 'Carlini & Wagner s L2 attack' and uses an 'Adam optimizer' and 'dropout', but does not provide specific software dependency names with version numbers (e.g., TensorFlow 2.x, PyTorch 1.x).
Experiment Setup Yes In our OPTMARGIN attack, we create a surrogate model of the region classifier, which classifies a smaller number of perturbed input points. We use 20 classifiers in the attacker s ensemble... We train with an Adam optimizer with a batch size of 128 and a learning rate of 0.001. For MNIST, we train on 8,000 examples (each example here contains both a benign image and an adversarial image) for 32 epochs, and we test on 2,000 other examples.