Lethal Dose Conjecture on Data Poisoning
Authors: Wenxiao Wang, Alexander Levine, Soheil Feizi
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, as a proof of concept, we show that by simply using different data augmentations for base learners, we can respectively double and triple the certified robustness of DPA on CIFAR-10 and GTSRB without sacrificing accuracy. (Abstract) and We prove the conjecture in multiples cases including Isotropic Gaussian classifications (Contributions) |
| Researcher Affiliation | Academia | Wenxiao Wang, Alexander Levine and Soheil Feizi Department of Computer Science University of Maryland College Park, MD 20742 {wwx, alevine0, sfeizi}@umd.edu |
| Pseudocode | No | No pseudocode or clearly labeled algorithm blocks are present in the paper. |
| Open Source Code | No | The paper's checklist indicates 'Yes' for providing code (3a), but there is no specific URL, repository link, or explicit statement within the main text or abstract indicating where the source code for the methodology described is available. It mentions 'supplemental material' but doesn't provide access details in the main paper. |
| Open Datasets | Yes | We use Network-In-Network[21] architecture for all base learners and evaluate on CIFAR-10[18] and GTSRB[29]. (Section 7.2) and Both CIFAR-10 and GTSRB are standard benchmarks. (Checklist 4e) |
| Dataset Splits | Yes | First, we evaluate the test accuracy of base learners with limited data (i.e. using 1/k of the entire training set of CIFAR-10 and GTSRB where k ranges from 50 to 500). (Section 7.2) and We use an initial learning rate of 0.005, a batch size of 16 for 600 epochs on CIFAR-10 and 1000 epochs on GTSRB. (Section 7.2) |
| Hardware Specification | No | Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] (Checklist 3d) |
| Software Dependencies | Yes | all included in torchvision[24]. (Section 7.2) and PyTorch: an imperative style, high-performance deep learning library. (Reference [24]) |
| Experiment Setup | Yes | For the baseline (DPA_baseline), we follow exactly their augmentations, learning rates (initially 0.1, decayed by a factor of 1/5 at 30%, 60%, and 80% of the training process), batch size (128) and total epochs (200). For our results (DPA_aug0 and DPA_aug1 on CIFAR-10; DPA_aug on GTSRB), we use predefined Auto Augment [6] policies for data augmentations... We use an initial learning rate of 0.005, a batch size of 16 for 600 epochs on CIFAR-10 and 1000 epochs on GTSRB. |