Towards Reverse-Engineering Black-Box Neural Networks
Authors: Seong Joon Oh, Max Augustin, Mario Fritz, Bernt Schiele
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We have introduced a procedure for constructing a dataset of classifiers (MNIST-NETS) as well as novel metamodels (kennen variants) that learn to extract information from black-box classifiers. In this section, we evaluate the ability of kennen to extract information from black-box MNIST digit classifiers. |
| Researcher Affiliation | Academia | Seong Joon Oh, Max Augustin, Bernt Schiele, Mario Fritz Max-Planck Institute for Informatics, Saarland Informatics Campus, Saarbr ucken, Germany {joon,maxaug,schiele,mfritz}@mpi-inf.mpg.de |
| Pseudocode | No | The paper describes algorithms using text and mathematical formulations but does not include any structured pseudocode or algorithm blocks (e.g., a figure labeled 'Algorithm 1'). |
| Open Source Code | Yes | The code is available at goo.gl/Mb Yfsv. |
| Open Datasets | Yes | We explain how MNIST-NETS has been constructed, a dataset of 11k MNIST digit classifiers; the procedure is task and data generic. and We also consider training MNIST classifiers on either on the entire MNIST training set (All0, 60k)... and We have resorted to 19 Py Torch3 pretrained Image Net classifiers. |
| Dataset Splits | Yes | Unless stated otherwise, every split has 5, 000 training (meta-training), 1, 000 testing (black box), and 5, 282 leftover models. |
| Hardware Specification | Yes | It takes around 5 minutes to train each model on a GPU machine (Ge Force GTX TITAN); training of 10k classifiers has taken 40 GPU days. |
| Software Dependencies | No | The paper mentions 'Py Torch3' but does not provide specific version numbers for PyTorch or any other software dependencies used in the experiments. |
| Experiment Setup | Yes | All the models have been trained with learning rate 0.1 and momentum 0.5 for 100 epochs. and We model the classifier mθ via multilayer perceptron (MLP) with two hidden layers with 1000 hidden units. |