Preemptive Image Robustification for Protecting Users against Man-in-the-Middle Adversarial Attacks

Authors: Seungyong Moon, Gaon An, Hyun Oh Song7823-7830

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on CIFAR-10 and Image Net show our method can effectively robustify natural images within the given modification budget.
Researcher Affiliation Collaboration 1 Department of Computer Science and Engineering, Seoul National University, Seoul, Korea 2 Deep Metrics, Seoul, Korea
Pseudocode Yes Algorithm 1: Preemptive robustification algorithm
Open Source Code Yes The code is available online 1. https://github.com/snu-mllab/preemptive_robustification
Open Datasets Yes We evaluate our methods on CIFAR-10 and Image Net by measuring classification accuracies of preemptively robustified images under the grey-box and white-box adversaries.
Dataset Splits No The paper mentions evaluating on CIFAR-10 and ImageNet, which are standard benchmarks, but does not explicitly provide specific training/validation/test split percentages, absolute sample counts for each split, or reference predefined splits with explicit citations for the validation set in the main text.
Hardware Specification No No specific hardware (GPU models, CPU models, or cloud computing instances with specifications) used for running experiments is explicitly mentioned in the paper.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al. 2019)' but does not provide specific version numbers for PyTorch or any other software dependencies needed for replication.
Experiment Setup Yes As it is natural to assume that the defender and the adversary have the same modification budget, we set δ = ϵ for all experiments. Both the adversaries use 20-step untargeted PGD and Auto Attack (Croce and Hein 2020) to find adversarial examples. For the white-box adversary, we sweep the final perturbation budget ϵ and report the lowest accuracy measured. We set the noise level to σ = 0.1. The noise levels are σ = 0.25 for CIFAR-10 and σ = 1.0 for Image Net.