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 [1].
Adversarial Reprogramming Revisited
Authors: Matthias Englert, Ranko Lazic
NeurIPS 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We initiate a theoretical study of adversarial reprogramming. In the experimental part of our work, we demonstrate that, as long as batch normalisation layers are suitably initialised, even untrained networks with random weights are susceptible to adversarial reprogramming. |
| Researcher Affiliation | Academia | Matthias Englert University of Warwick EMAIL Ranko Lazi c University of Warwick EMAIL |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We are making code to run the experiments available at https://github.com/englert-m/adversarial_reprogramming. |
| Open Datasets | Yes | We use the same dataset, which consists of 60,000 training images and 10,000 test images, for our experiments. It is available under the Creative Commons Attribution Share-Alike 3.0 licence. |
| Dataset Splits | No | The paper states the use of 60,000 training images and 10,000 test images, but does not explicitly mention a separate validation set split for hyperparameter tuning or early stopping during their experiments. |
| Hardware Specification | Yes | The experiments were mainly run on two internal clusters utilising a mix of NVIDIA GPUs such as Ge Force RTX 3080 Ti, Quadro RTX 6000, Ge Force RTX 3060, Ge Force RTX 2080 Ti, and Ge Force GTX 1080. |
| Software Dependencies | Yes | We use the networks exactly as implemented in Keras in Tensor Flow 2.8.1 |
| Experiment Setup | Yes | We use the 60,000 training images to run an Adam optimiser [Kingma and Ba, 2015] with learning rate 0.01 and a batch size of 50 to optimise the unconstrained weights of the adversarial program. |