Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

GSE: Group-wise Sparse and Explainable Adversarial Attacks

Authors: Shpresim Sadiku, Moritz Wagner, Sebastian Pokutta

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Rigorous evaluations on CIFAR-10 and Image Net datasets demonstrate a remarkable increase in groupwise sparsity, e.g., 50.9% on CIFAR-10 and 38.4% on Image Net (average case, targeted attack). This performance improvement is accompanied by significantly faster computation times, improved explainability, and a 100% attack success rate.
Researcher Affiliation Academia Shpresim Sadiku1,2, Moritz Wagner1,2 & Sebastian Pokutta1,2 1Department for AI in Society, Science, and Technology, Zuse Institute Berlin, Germany 2Institute of Mathematics, Technische Universität Berlin, Germany EMAIL
Pseudocode Yes Algorithm 1 Forward-Backward Splitting Attack
Open Source Code Yes All tests are conducted on a machine with an NVIDIA A40 GPU, and our codes, 10k image indices from the Image Net validation dataset, and target labels for targeted Image Net tests are available at https://github.com/wagnermoritz/GSE.
Open Datasets Yes We experiment on CIFAR-10 (Krizhevsky et al., 2009) and Image Net (Deng et al., 2009) datasets, analyzing DNNs on 10k randomly selected images from both validation sets.
Dataset Splits Yes We experiment on CIFAR-10 (Krizhevsky et al., 2009) and Image Net (Deng et al., 2009) datasets, analyzing DNNs on 10k randomly selected images from both validation sets. For the classifier C on CIFAR-10, we train a Res Net20 model (He et al., 2016) for 600 epochs using stochastic gradient descent, with an initial learning rate of 0.01, reduced by a factor of 10 after 100, 250, and 500 epochs. We set the weight decay to 10 4, momentum to 0.9, and batch size to 512. For Image Net, we employ a Res Net50 (He et al., 2016) and a more robust transformer model, Vi T_B_16 (Dosovitskiy et al., 2020), both with default weights from the torchvision library.
Hardware Specification Yes All tests are conducted on a machine with an NVIDIA A40 GPU
Software Dependencies No The paper mentions "Py Torch" in Section 3.3, but does not specify a version number.
Experiment Setup Yes For the classifier C on CIFAR-10, we train a Res Net20 model (He et al., 2016) for 600 epochs using stochastic gradient descent, with an initial learning rate of 0.01, reduced by a factor of 10 after 100, 250, and 500 epochs. We set the weight decay to 10 4, momentum to 0.9, and batch size to 512. Specifically, for CIFAR-10, we set q = 0.25, σ = 0.005, µ = 1, and ˆk = 30, while for Image Net, we use q = 0.9, σ = 0.05, µ = 0.1, and ˆk = 50. We run all the attacks for a total of 200 iterations.