Robust and Privacy-preserving AI in Crowdsensing Platforms

Content 🏷️

This repository contains all the code and data required to run the experiments, respectively recreate the figures of the report of the corresponding master project. It is implemented using Python 3 and additional libraries. The DL models are built using PyTorch. The results of this project have been condensed into a short paper published by the JOURNAL IEEE Transactions on Dependable and Secure Computing. The publication can be found in IEEEExplore.

Setup 🔨

In order to run the code, Python 3.8+ must be used. Subsequently, all the required libraries (listed in the requirements.txt file) must be installed. It is recommended to set up a virtual environment (e. g. using venv) and then install the libraries there. Additionally, PyTorch must be installed using the installation instructions provided on their website.

GPU Support ⚡

GPU Support is available in PyTorch. The code is set up to run on the CPU, as most experiments run within a reasonable time on CPU. In order to run them on GPU, the instructions provided by PyTorch need to be followed: CUDA must be setup properly and all the variables and models must be put on the GPU, e. g. using the built-in .cuda() command.

Code structure 📖

• /data/: The raw CSVs of the monitoring. The subdirectories are mapped to the device ids as follows:

  • ras-3-1gb: pi3-1gb
  • ras-4-4gb: pi4-4gb
  • ras-4-black: pi4-2b-bc
  • ras-4-white: pi4-2gb-wc

• /data_exploration/: Runnables to generate the visualizations corresponding to the data exploration from Section 3.3

• /federated_anomaly/: Runnables for the experiments from Section 5.3

• /federated_anomaly_detection/: Runnables for the experiments from Section 6.2

• /federated_binary_classification/: Runnables for the experiments from Section 5.4

• /federated_binary_classification_adversarial/: Runnables for the experiments from Section 6.3

• /standalone/: Runnables for the experiments from Chapter 4

• /various/: configurable experiments used for testing and validation, not included in report.

• aggregation.py: Implements the different Aggregation Functions as well as Threshold Selection for a Federation on top of a list of Participants: Specifically those are:

  • FedAvg
  • Trimmed Mean 1 and 2
  • Coordinate Wise Median

• custom_types.py: Implements the specific ENUM classes for this project: Device Types (RaspberryPi), Behaviors etc.

• data_handler.py: Implements the dataset splitting and scaling for a given definiton of Participants: Generates train, validation and test sets for them and can further be used to scale them

• models.py: Implements the PyTorch Sequential models of the MLP as well as the Autoencoder

• participants.py: Implements a Participant, which can then either be an AutoEncoderParticipant (for anomaly detection) or an MLPParticipant (for binary classification). Further implements adversarial participants as subclasses.

• utils.py: Implements common utility functions for multiple Experiments. Includes reporting utilities as well as definition of (balanced) federation participants