Detecting Adversarial Examples Through Image Transformation

Authors: Shixin Tian, Guolei Yang, Ying Cai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments with two datasets show that our technique can detect nearly 99% of adversarial examples generated by the state-of-the-art algorithm.
Researcher Affiliation Academia Shixin Tian, Guolei Yang, Ying Cai Department of Computer Science, Iowa State University {stian,yanggl,yingcai}@iastate.edu
Pseudocode No The paper describes the method in text and a diagram (Figure 5) but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information (e.g., repository link, explicit statement of code release) for the source code of their methodology.
Open Datasets Yes The MNIST dataset has 70,000 handwritten digits from which 60,000 are used as the training set and 10,000 as the testing set. The CIFAR10 dataset consists of 60,000 colour images in 10 classes. 50,000 of them are used as the training set and the rest as the testing set.
Dataset Splits No The paper explicitly mentions training and testing sets for the main classifiers and the detector, but does not provide details on a specific validation set split.
Hardware Specification Yes When performing our experiments on a laptop with GPU (NVIDIA Ge Force GTX 960M), generating one adversarial example for oblivious threat model takes around 4 seconds while generating one adversarial example for white-box threat model takes around 75 seconds in average.
Software Dependencies No The paper states "The experiments are implemented with Keras using Tensor Flow as backend" but does not provide specific version numbers for these software components.
Experiment Setup Yes These classifiers are trained at the learning rate of 0.01 with a batch size of 128 and 50 epochs.