Markpainting: Adversarial Machine Learning meets Inpainting
Authors: David Khachaturov, Ilia Shumailov, Yiren Zhao, Nicolas Papernot, Ross Anderson
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the performance of targeted markpainting in Section 5.2, and the effect of different masks and target images in Section 5.3. Section 5.4 focuses on the transferability of the generated samples, while Section 5.5 discusses mask-agnostic markpainting. and Table 2. Impact of markpainting on different inpainter models. This table reports the loss (Lmark from Section 4.2), L2 norms, peak signal to noise ratio (PSNR) and structural index similarity (SSIM) for assessing the inpainted image quality. |
| Researcher Affiliation | Academia | 1Computer Laboratory, University of Cambridge 2University of Toronto and Vector Institute. |
| Pseudocode | Yes | Algorithm 1 General markpainting algorithm and Algorithm 2 Eo T markpainting algorithm |
| Open Source Code | Yes | Source code is available at: https://github. com/iliaishacked/markpainting. |
| Open Datasets | Yes | Table 1. Inpainters used in the evaluation System Dataset Generative (Yu et al., 2018) Image Net (Deng et al., 2009) GMCNN (Wang et al., 2018) Celeb A-HQ (Liu et al., 2015) Edge Connect (Nazeri et al., 2019) Paris Street View (Doersch et al., 2012), Celeb A (Liu et al., 2015), Places2 (Zhou et al., 2017) RFR (Li et al., 2020b) Paris Street View, Celeb A RN (Yu et al., 2020) Places2 CRFILL (Zeng et al., 2020) Places2, Salient Object Segmentation (Xiong et al., 2019) |
| Dataset Splits | No | The paper does not explicitly provide training, validation, and test dataset splits that they used for their experiments, only mentions evaluation on a subset of a dataset and using pretrained models. |
| Hardware Specification | No | The paper does not specify the hardware (e.g., GPU, CPU models) used for conducting the experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with their version numbers. |
| Experiment Setup | Yes | A maximum perturbation budget of ϵ was used with a step size of ϵ = ϵ 50 unless specified otherwise. We justify the parameter choices in Section 4 of the Appendix. We clip the markpainted image at each iteration to make sure that the total perturbation budget does not exceed ϵ. and We set α = 4, based on experimentation. and The number of iterations was taken to be 1500 with a step size of ϵ 30, with mmin = 0.01 and mmax = 0.1. |