Black-Box Ripper: Copying black-box models using generative evolutionary algorithms

Authors: Antonio Barbalau, Adrian Cosma, Radu Tudor Ionescu, Marius Popescu

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on three benchmark data sets: CIFAR-10 [18] with CIFAR-100 [18] as proxy, Fashion-MNIST [37] with CIFAR-10 [18] as proxy and 10 Monkey Species [27] with Celeb A-HQ [13] and Image Net Cats and Dogs [33] as proxies.
Researcher Affiliation Academia Antonio B arb al au1, , Adrian Cosma2, Radu Tudor Ionescu1, Marius Popescu1 1University of Bucharest, Romania 2University Politehnica of Bucharest, Romania
Pseudocode Yes Algorithm 1 Evolutionary Optimization Algorithm
Open Source Code Yes Our code is available at: https://github.com/antoniobarbalau/black-box-ripper.
Open Datasets Yes We conduct experiments on three benchmark data sets: CIFAR-10 [18] with CIFAR-100 [18] as proxy, Fashion-MNIST [37] with CIFAR-10 [18] as proxy and 10 Monkey Species [27] with Celeb A-HQ [13] and Image Net Cats and Dogs [33] as proxies.
Dataset Splits No The paper defines 'original (or true) training set', 'original (or true) test set', and 'proxy training set', but does not specify details regarding a validation dataset split or its proportion.
Hardware Specification No The paper describes the neural network architectures used (e.g., Alex Net, VGG-16, Le Net, Res Net-18) but does not provide specific details about the hardware, such as GPU or CPU models, used for running the experiments.
Software Dependencies No The paper mentions various models and optimizers like VAE [16], GAN [7], Adam [15], Progressively Growing GAN (Pro GAN) [13], Spectral Normalization GAN (SNGAN) [26], but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes All models, including baselines, are trained for 200 epochs using the Adam [15] optimizer. We used mini-batches of 64 images. In Black-Box Ripper, the images are synthesized by the evolutionary algorithm, using at most 10 iterations, a population of K = 30 latent vectors sampled within the boundary u = 3, an elite size of k = 10, halting the optimization if the fittest latent vector gives an objective value lower than t = 0.02.