Final Project CMPUT 644

Detecting Adversarial Attacks on Federated Learning with Unbalanced Classes

Intro

In this project we investigate the resilience of proposed poisoning attacks and detection systems in a variety of adversarial proposed in [1]. Our main contributions, in addition to reproducing the poisoning attacks and corresponding detection strategies are as follows:

• We proposed a smooth and incremental method to generate unbalanced datasets and evaluated the performance of FedAvg, and the proposed poisoning detection mechanisms on unbalanced datasets.
• We implemented adversaries with varying amounts of poison, and model the behavior of the proposed poisoning detection strategies.
• We conducted experiments with two distinct distance metrics for one of the detection algorithms, and for the first time presented experimental results on the performance of detection metrics.

[1] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. 2019. Analyzing federated learning through an ad- versarial lens. In International Conference on Machine Learning. PMLR, 634–643.

Experiments:

1. Reproducing Malicious attacks: flower_args.py
2. Effect of Malicious Attack on Model Accuracy: flower_args.py
3. Detecting the stealthy malicious attack: The weights of flower_args.py are saved (directory not uploaded to repo as it is ~51GB) 3.1 Validation accuracy-based detection: utils/detection_algo_val_accuracy 3.2 Distance ranges-based detection: DetectionStats/Step2_DataCrunching.ipynb, DetectionStats/Step3_Visualization.ipynb
4. Accuracy versus class imbalance: run_imbalance_v_accuracy.py
5. Detecting stealthy attacks on unbalanced data: flower_args.py is run with different parameters. See $python3 flower_args.py --help 5.1 Validation accuracy-based detection: utils/detection_algo_val_accuracy 5.2 Distance ranges-based detection: DetectionStats/Step2_DataCrunching.ipynb, DetectionStats/Step3_Visualization.ipynb
6. Summary of threshold values: DetectuibStats/Step4_SummarizeKappa.ipynb

Dependencies:

The dependencies can be checked on requirements.txt. Using a venv is recommended.

Acknowledgments:

We'd like to thank our Instructor and Supervisor for his wonderful teachings and guidance during this course.

Running on Snowball notes:

To use GPU 1 or 2, instead of 0 on snowball. Do export CUDA_VISIBLE_DEVICES=1,2 on the command line before running python script.