DisGUIDE: Disagreement-Guided Data-Free Model Extraction
Authors: Jonathan Rosenthal, Eric Enouen, Hung Viet Pham, Lin Tan
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our evaluation on popular datasets CIFAR-10 and CIFAR-100 shows that our approach improves the final model accuracy by up to 3.42% and 18.48% respectively. The average number of queries required to achieve the accuracy of the prior state of the art is reduced by up to 64.95%. |
| Researcher Affiliation | Academia | Jonathan Rosenthal1, Eric Enouen2*, Hung Viet Pham3 , Lin Tan1 1 Purdue University 2 The Ohio State University 3 York University |
| Pseudocode | No | The paper describes the Dis GUIDE training process but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Dis GUIDE codebase: https://github.com/lin-tan/disguide |
| Open Datasets | Yes | We evaluate Dis GUIDE on the two widely-used image classification datasets CIFAR-10 and CIFAR-100 (Krizhevsky and Hinton 2009). |
| Dataset Splits | No | The paper refers to 'held-out test sets' and 'training data' but does not provide specific train/validation/test split percentages, sample counts for each split, or explicit methodology for creating these splits. It uses well-known datasets like CIFAR-10 and CIFAR-100 which have standard splits, but these are not explicitly detailed within the paper's text. |
| Hardware Specification | Yes | We conduct our experiments on a server with 48 CPU cores with 504 GB of RAM and 2080Ti GPUs. |
| Software Dependencies | Yes | Our code uses Pytorch 1.11 and CUDA 10.2. |
| Experiment Setup | Yes | We use the same generator training hyperparameters as DFME: a batch size of 256, Adam optimizer with an initial learning rate of 1e-4 and weight decay of 5e-4. Similarly, we use DFME s hyperparameters for clone training: batch size of 256, SGD with an initial learning rate of 0.1, and the same weight decay as above. ... We select b = 3 as well as a replay buffer size s = 1M ... We empirically set the class diversity loss weight λ = 0.2 and λ = 0.04 for CIFAR-10 and CIFAR-100 experiments respectively. ... We empirically select 1/8 of the generated samples to be set to grayscale. |