Enhancing Robustness of Neural Networks through Fourier Stabilization

Authors: Netanel Raviv, Aidan Kelley, Minzhe Guo, Yevgeniy Vorobeychik

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally evaluate the proposed Fourier stabilization approach on several datasets involving detection of malicious inputs, including malware detection and hate speech detection. Our experiments show that our approach considerably improves neural network robustness to evasion in these domains, and effectively composes with adversarial training defense.
Researcher Affiliation Academia Netanel Raviv 1 Aidan Kelley 1 Michael Guo 1 Yevgeny Vorobeychik 1 1Department of Computer Science and Engineering, Washington University in St. Louis, 1 Brookings Dr., St. Louis, MO 63103.
Pseudocode No The paper describes algorithmic approaches (GMBC, GMB) in Section 4 but does not present them in a structured pseudocode or algorithm block.
Open Source Code Yes Experiments were run on a research computer cluster with over 2,500 CPUs and 60 GPUs, and the code is available online at https://github.com/ Aidan Kelley/fourier-stabilization.
Open Datasets Yes The PDFRate dataset (Smutz & Stavrou, 2012) is a PDF malware dataset... The Hidost dataset (Srndic & Laskov, 2016) is a PDF malware dataset... The Hate Speech dataset (Qian et al., 2019), collected from Gab, is comprised of conversation segments...
Dataset Splits No All datasets were divided into training, validation, and test subsets; the former two were used for training and parameter tuning, while all the results below are using the test data. The paper states that data was divided into these subsets but does not provide specific percentages or counts for reproducibility.
Hardware Specification No Experiments were run on a research computer cluster with over 2,500 CPUs and 60 GPUs. This provides the quantity of hardware but not specific models or detailed specifications for reproducibility.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes For each dataset, we learned a two-layer sigmoidal fully connected neural network as a baseline.