Generalised Lipschitz Regularisation Equals Distributional Robustness

Authors: Zac Cranko, Zhan Shi, Xinhua Zhang, Richard Nock, Simon Kornblith

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental As our experiments show, this method achieves higher robustness than state of the art (Cisse et al., 2017; Anil et al., 2019). We studied the empirical robustness and accuracy of the proposed Lipschitz regularisation technique for adversarial training of kernel methods, under both Gaussian kernel and inverse kernel.
Researcher Affiliation Collaboration 1Universität Tübingen, Tübingen, Germany 2University of Illinois at Chicago, IL, USA 3Google Brain.
Pseudocode Yes Algorithm 1 Training L-Lipschitz binary SVM
Open Source Code No The paper references 'LNets. https://github.com/cemanil/LNets.' in the bibliography, which is a third-party tool used for comparison. However, the authors do not state that they are releasing their own source code for the methodology described in this paper.
Open Datasets Yes Datasets We tested on three datasets: MNIST, Fashion-MNIST, and CIFAR10.
Dataset Splits Yes The number of training/validation/test examples for the three datasets are 54k/6k/10k, 54k/6k/10k, 45k/5k/10k, respectively.
Hardware Specification No The paper does not specify the exact hardware (e.g., GPU models, CPU types, memory) used for running the experiments. It only implies computations were performed.
Software Dependencies No The paper mentions using PGD and L-BFGS, but it does not specify any software names with version numbers (e.g., specific library versions for PyTorch, TensorFlow, or scikit-learn).
Experiment Setup Yes The perturbation δ was constrained in an 2-norm or 1-norm ball. To evaluate robustness, we scaled the perturbation bound δ from 0.1 to 0.6 for 1-norm norm, and from 1 to 6 for 2-norm norm (when δ = 6, the average magnitude per coordinate is 0.214). We normalised gradient and fine-tuned the step size. To defend against 2-norm attacks, we set L = 100 for all algorithms. Gauss-Lip achieved high accuracy and robustness on the validation set with bandwidth σ = 1.5 for Fashion MNIST and CIFAR-10, and σ = 2 for MNIST. To defend against 1-norm attacks, we set L = 1000 for all the four methods as in Anil et al. (2019). The best σ for Gauss-Lip is 1 for all datasets. Inverse-Lip used 5 layers.