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].
Fingerprints of Super Resolution Networks
Authors: Jeremy Vonderfecht, Feng Liu
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To understand the fingerprints of SISR networks, we collect photographs from Flickr and super-resolve each of them with 205 different SISR models. These 205 models consist of 25 pretrained models published online by other researchers, and 180 models which we have trained ourselves by systematically varying four experimental hyperparameters: architecture, super-resolution scale, training dataset, and loss function. We then train an extensive collection of image classifiers to perform model attribution, and to predict the values of our experimental hyperparameters. By systematically reserving different subsets of the SISR models for testing, we investigate how our model attribution and parsing classifiers generalize. |
| Researcher Affiliation | Academia | Jeremy Vonderfecht EMAIL Department of Computer Science Portland State University, Feng Liu EMAIL Department of Computer Science Portland State University |
| Pseudocode | No | The paper describes methods in prose and refers to figures, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | The paper mentions that 'Our 180 custom-trained SISR models were trained using the Basic SR framework (Wang et al., 2018a)' and cites 'Basic SR: Open source image and video restoration toolbox. https://github.com/xinntao/Basic SR, 2018a.'. This refers to a third-party tool used by the authors, not the open-sourcing of the authors' own methodology or implementation for their fingerprint analysis and parsing classifiers. |
| Open Datasets | Yes | We develop a novel dataset of 205,000 super-resolved images generated by 205 different SISR models, all of which will be made publicly available. |
| Dataset Splits | Yes | Our classifiers are trained with our super-resolved image dataset on just 800 images from each SISR model. We reserve an additional 100 images for validation, and 100 for testing. Images are cropped down to the network s input size. In the training set, images are randomly cropped. In validation and testing, they are center-cropped. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used, such as GPU models, CPU specifications, or memory. |
| Software Dependencies | No | The paper mentions using a 'Conv Next network', 'Adamax optimizer', and the 'Basic SR framework (Wang et al., 2018a)', but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | The network is trained with cross-entropy loss. As in Rössler et al. (2019), we train just the final layer for three epochs, then we unfreeze all network weights and fine-tune the entire network for 15 epochs. We use the Adamax optimizer with a learning rate of 0.0005, Batch size of 16. Our classifiers are trained with our super-resolved image dataset on just 800 images from each SISR model. We reserve an additional 100 images for validation, and 100 for testing. ... all models were trained for 50,000 iterations. |