Attacks on Online Learners

This repository contains the code implementing the attacks on data labels presented in the paper 'Attacks On Online Learners: a Teacher-Student Analysis' (https://arxiv.org/abs/2305.11132).

• Real data greedy experiments

  Perceptron, MNIST:

  • Script: Experiment_PerceptronMNIST - Input: index (int) for the C-grid.
  • Imports parameters from ParametersPerceptronMNIST.py.
  • Runs several experiments (n_runs_experiments) for the chosen parameters.
  • Uses opt_pref=True, meaning that each simulation first runs a calibration of the future weight pre-factor ($\tilde{\gamma}$).

  LeNet, MNIST:

  • Script: Experiment_LeNetMNIST.py - Input: index (int) for the future weight pre-factor ($\tilde{\gamma}$) grid.
  • Imports parameters from ParametersLeNetMNIST.py.
  • Runs several experiments (n_runs_experiments) for the chosen parameters.
  • Uses opt_pref=False - the experiment reaching the lowest steady-state running average is used for downstream analysis.

  VGG11, CIFAR10 (transfer learning):

  • First train a teacher function using TeacherTraining_VGG11TransferCIFAR10.py.
  • Note that we use pre-trained weights imported from Torchvision and only train the last layer.
  • Then use Export_VGG11DataCIFAR10.py to export pre-last layer activations.
  • Finally, use the exported data to run experiments via Experiment_TransferVGG11CIFAR10.py.

  ResNet18, CIFAR10 (transfer learning):

  • Experiments follow the same protocol used for VGG11.

• Synthetic data experiments

  Greedy attacks:

  • Linear regression. Scripts: Experiment_GreedyAttacks_LinearRegression.py, Experiment_GreedyAttacks_LinearRegression_FractionPoisonedSamples.py, parameters loaded from ParametersGreedyAttacks_LinearRegression.py. Notebook: Experiment_GreedyAttacks_LinearRegression_MultiDimControl.ipynb.
  • Sigmoidal perceptron: Script Experiment_GreedyAttacks_SigmoidalPerceptron.py, parameters loaded from ParametersGreedyAttacks_SigmoidalPerceptron.py.
  • 2-layer NN. Script: Experiment_GreedyAttacks_2LayerNN.py, parameters loaded from ParametersGreedyAttacks_2LayerNN.py.

  Strategies comparison:

  • For each architecture (linear regression, sigmoidal perceptron, 2-layer NN):
  • Parameters are set in the corresponding file ParametersAttacksComparison_*.py.
  • Train the RL TD3 agent using the notebook RLAgentTraining_*.ipynb.
  • Compare constant, reinforcement learning, greedy, and clairvoyant attacks using the notebook StrategiesComparison_*.ipynb.