/ST-Siamese-Attack

CW-attack and FGSM-attack on ST-Siamese network

Primary LanguageJupyter Notebook

ST-iFGSM: Enhancing Robustness of Human Mobility Signature Identification Model via Spatial-Temporal Iterative FGSM

This repository contains a Keras implementation of the algorithm presented in the paper ST-iFGSM: Enhancing Robustness of Human Mobility Signature Identification Model via Spatial-Temporal Iterative FGSM

In this paper, we aim to improve the robustness of the HuMID model by generating useful adversarial trajectories for further training the model. To accomplish this, we design a Spatial Temporal iterative Fast Gradient Sign Method with 𝐿0 regularization – ST-iFGSM – to generate adversarial attacks on state-of-the-art (SOTA) HuMID models.

The solution framework takes the human mobility dataset and a target HuMID model as inputs and contains two stages: Stage 1. iteratively generating and selecting adversarial attack samples which could fool the target HuMID model, and Stage 2. training the HuMID model with the adversarial attack to improve the model robustness.

ST-iFGSM L0 framework

Prerequisites

Usage

  • Execute python classification.py to train the multi-classification model.
  • Execute python train.py to train ST-siamese model.
  • Execute python classification_attack.py to attack the multi-classification model
  • Execute python siamese_attack.py to attack the ST-Siamese model
  • Execute python fast_adversarial_train.py to do Fast ST-FGSM adversarial train