Thwarting Adversarial Examples: An $L_0$-Robust Sparse Fourier Transform

Authors: Mitali Bafna, Jack Murtagh, Nikhil Vyas

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We give experimental results on the Jacobian-based Saliency Map Attack (JSMA) and the Carlini Wagner (CW) L0 attack on the MNIST and Fashion-MNIST datasets as well as the Adversarial Patch on the Image Net dataset.
Researcher Affiliation Academia Mitali Bafna School of Engineering & Applied Sciences Harvard University Cambridge, MA USA mitalibafna@g.harvard.edu Jack Murtagh School of Engineering & Applied Sciences Harvard University Cambridge, MA USA jmurtagh@g.harvard.edu Nikhil Vyas Department of Electrical Engineering and Computer Science MIT Cambridge, MA USA nikhilv@mit.edu
Pseudocode Yes Algorithm 1 Iterative Hard Thresholding (IHT) [BCDH10].
Open Source Code No The paper does not contain any statement about releasing source code or provide links to a code repository.
Open Datasets Yes We tested both JSMA and CW on two datasets: the MNIST handwritten digits [Le C98] and the Fashion-MNIST [XRV17] dataset of clothing images. ... We took 700 random images from Image Net and for classification we used pretrained Res Net-50 network [HZRS15].
Dataset Splits No The paper mentions 'training datasets' and reports 'test accuracy', but it does not specify explicit training/validation/test splits, percentages, or sample counts for these splits.
Hardware Specification No The paper does not specify any particular hardware (e.g., CPU, GPU models, or cloud resources) used for running the experiments.
Software Dependencies No The paper mentions using the 'Adam optimizer' and 'cross-entropy loss' but does not provide specific version numbers for any software, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For both datasets we used a neural network composed of a convolutional layer (32 kernels of 3x3), max pooling layer (2x2), convolutional layer (64 kernels of 3x3), max pooling layer (2x2), fully connected layer (128 neurons) with dropout (rate = .25) and an output softmax layer (10 neurons). We used the Adam optimizer with cross-entropy loss and ran it for 10 epochs over the training datasets. For each dataset, we trained our neural network only on images that were projected onto their top-k 2D-DCT coefficients. Here k is a parameter we tuned depending on the dataset (for MNIST k = 40 and for Fashion-MNIST k = 35).