PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks
Authors: Jan Svoboda, Jonathan Masci, Federico Monti, Michael Bronstein, Leonidas Guibas
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper we introduce Peer Nets, a novel family of convolutional networks alternating classical Euclidean convolutions with graph convolutions to harness information from a graph of peer samples. This results in a form of non-local forward propagation in the model, where latent features are conditioned on the global structure induced by the graph, that is up to 3 more robust to a variety of whiteand black-box adversarial attacks compared to conventional architectures with almost no drop in accuracy. |
| Researcher Affiliation | Collaboration | Jan Svoboda1,2, Jonathan Masci2, Federico Monti1,4, Michael M. Bronstein1,3,4,5, Leonidas Guibas6 1USI, Switzerland 2NNAISENSE, Switzerland 3Intel Perceptual Computing, Israel 4Imperial College London, UK 5Fabula AI, UK 6Stanford University, USA {jan.svoboda,federico.monti,michael.bronstein}@usi.ch jonathan@nnaisense.com, guibas@cs.stanford.edu |
| Pseudocode | No | The paper describes the Peer Regularization layer with mathematical formulas and textual descriptions but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper does not explicitly state that the source code for the Peer Nets methodology is available, nor does it provide a direct link to the authors' own repository. It does link to third-party code for baseline comparisons in Appendix D and E, but these are not the authors' own implementations for their method. |
| Open Datasets | Yes | We evaluate the robustness of Peer Regularization to adversarial perturbations on standard benchmarks (MNIST Le Cun (1998), CIFAR-10, and CIFAR-100 Krizhevsky (2009)) |
| Dataset Splits | No | The paper mentions using standard datasets and describes training and testing procedures, but it does not explicitly provide details about specific validation dataset splits (e.g., percentages, sample counts, or explicit reference to a standard validation split) required for reproduction. |
| Hardware Specification | No | The paper mentions that 'due to the limited GPU memory, we typically had to decrease the batch size for the PR network architectures,' implying the use of GPUs. However, it does not specify any particular GPU model (e.g., NVIDIA V100, RTX 3090) or any other detailed hardware specifications like CPU model or memory. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software libraries, frameworks, or programming languages used in the experiments. It only references third-party tools by name without version details (e.g., 'Foolbox by Rauber et al. (2017)', 'Deep Fool toolbox'). |
| Experiment Setup | Yes | The training hyper-parameters used in our experiments are summarized in Table 1. Learning rate is decreased at epochs 100, 175, and 250 with a step factor of 10 1. Table 1: Optimization parameters for different architectures and datasets. Learning rate is decreased at epochs 100, 175, and 250 with a step factor of 10 1. Model Optimizer Epochs Batch Momentum LR L2 reg. LR decay Le Net-5 Adam 100 128 10 3 10 4 PR-Le Net-5 Adam 100 32 10 3 10 4 Res Net-32 Momentum 350 128 0.9 10 1 10 3 step PR-Res Net-32 Momentum 350 64 0.9 10 2 10 3 step Res Net-110 Momentum 350 128 0.9 10 1 2 10 3 step PR-Res Net-110 Momentum 350 64 0.9 10 2 2 10 3 step |