Certified Robustness against Sparse Adversarial Perturbations via Data Localization

Authors: Ambar Pal, Rene Vidal, Jeremias Sulam

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Empirical Evaluation In this section, we will briefly describe existing methods for probabilistic ℓ0 certification, (Levine & Feizi, 2020b) and (Jia et al., 2022) as well as deterministic ℓ0 certification (Hammoudeh & Lowd, 2023), and then empirically compare our (deterministic) ℓ0 certified defense Box-NN to these approaches.
Researcher Affiliation Academia Ambar Pal EMAIL Department of Computer Science & Mathematical Institute for Data Science Johns Hopkins University Baltimore, MD 21218, USA; René Vidal EMAIL Department of Electrical and Systems Engineering & Center for Innovation in Data Engineering and Science University of Pennsylvania Philadelphia, PA 19104, USA; Jeremias Sulam EMAIL Department of Biomedical Engineering & Mathematical Institute for Data Science Johns Hopkins University Baltimore, MD 21218, USA
Pseudocode No The paper describes the Box-NN classifier and methods for its development and certification (Section 4), including formulas for calculating distances (Lemma 4.1) and robustness certificates (Theorem 4.2). However, it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' block with structured, code-like steps.
Open Source Code No The paper does not contain any explicit statement about releasing source code, nor does it provide a link to a code repository or mention code in supplementary materials. The link provided 'https: // openreview. net/ forum? id= 17Ld3davz F' is for the OpenReview forum for the paper.
Open Datasets Yes We provide empirical evaluation on the MNIST and the Fashion-MNIST datasets, and demonstrate that Box-NN obtains state-of-the-art results in certified ℓ0 robustness.
Dataset Splits No For each of the methods described so far, we plot Cert Acc against ϵ using the corresponding robust classifier g and the certificate C over samples from the test set of the datasets mentioned. The paper does not explicitly provide details about the training, validation, or test dataset splits (e.g., percentages, sample counts, or specific splitting methodology).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No We ablate over a few choices of the gradient-based optimizer for our problem: (a) vanilla SGD with a learning rate of 0.02, (b) SGD with a learning rate of 0.02, a momentum of 0.9, and a weight decay of 0.0005, and (c) Adam with a learning rate of 0.001, and standard decay factors, in Fig. 8. The paper mentions optimizers but does not specify any software libraries or their version numbers.
Experiment Setup Yes We initialize θ by using a set of boxes defined from the data. This is done by first drawing a subset T of size M uniformly at random from the training data-points, and then initializing θ with axis-aligned boxes centered at these data-points, as {(B(x 0.1, x + 0.1), y): (x, y) T}... We clip the certificates to 50. ... We ablate over a few choices of the gradient-based optimizer for our problem: (a) vanilla SGD with a learning rate of 0.02, (b) SGD with a learning rate of 0.02, a momentum of 0.9, and a weight decay of 0.0005, and (c) Adam with a learning rate of 0.001, and standard decay factors.