Adversarial Feature Desensitization

Authors: Pouya Bashivan, Reza Bayat, Adam Ibrahim, Kartik Ahuja, Mojtaba Faramarzi, Touraj Laleh, Blake Richards, Irina Rish

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on several benchmarks demonstrate the effectiveness of the proposed approach against a wide range of attack types and attack strengths.
Researcher Affiliation Academia 1 Mc Gill University, Montreal, Canada 2 MILA, Université de Montréal, Montreal, Canada *Correspondence to: {bashivap,irina.rish}@mila.quebec
Pseudocode Yes Algorithm 1: AFD training procedure
Open Source Code Yes Our code is available at https://github.com/Bashivan Lab/afd.
Open Datasets Yes Datasets. We validated our proposed method on several common datasets including MNIST [30], CIFAR10, CIFAR100 [29], and tiny-Imagenet [26].
Dataset Splits Yes To find the best learning rates, we randomly split the CIFAR10 train set into a train and validation sets (45000 and 5000 images in train and validation sets respectively).
Hardware Specification Yes All experiments were run on NVIDIA V100 GPUs. We used one GPU for experiments on MNIST and 2 GPUs for other datasets.
Software Dependencies No The paper mentions using "Foolbox [42] and Advertorch [12] Python packages" but does not specify their version numbers or any other software dependencies with version information.
Experiment Setup Yes We used ϵ = 0.3, 0.031, and 0.016 for MNIST, CIFAR, and Tiny-Imagenet datasets respectively. ... Based on this analysis, we selected the learning rate γ = 0.5 for tuning the feature extractor Fθ, and α = β = 0.1 for tuning the parameters in domain discriminator Dψ, and the task classifier Cφ.