Independent Study Module (ISM): Secure ML

We studied different machine learning problems and ways to solve them in the first half of the ISM, and explored different data poisoning attacks that can be done on the ML algorithms.

ISM Final Report: Final Presentation.pdf

Implementation Details

Trash Bucket Problem:

• binary classfication: dog-vs-cat model created using CNN. Dataset used: https://www.kaggle.com/c/dogs-vs-cats/data https://www.kaggle.com/andrewmvd/animal-faces
• multiclass classfication: number recognition model created using simple NN. Dataset used: MNIST
• Notebooks: dogvscat.ipynb & MNIST_Trash_Bucket.ipynb

One shot learning with a Siamese Network:

• Paper: https://www.cs.cmu.edu/~rsalakhu/papers/oneshot1.pdf
• Dataset: MNIST & Omniglot, https://github.com/brendenlake/omniglot
• Notebooks: one_shot_learning_omniglot.ipynb & One_Shot_Learning_MNIST.ipynb

Subpopulation Data Poisoning Attack (Paper):

• Paper: https://arxiv.org/abs/2006.14026
• Dataset: Adult Dataset, https://archive.ics.uci.edu/ml/datasets/adult
• Notebooks: subpopulation_attack_paperimplementation.ipynb
• View Subpopulation Attack Paper Summary.pdf for the summary of the paper.

Subpopulation Data Poisinong Attack (real-world datasets):

• Dataset: Credit Risk (https://www.kaggle.com/laotse/credit-risk-dataset) & Stroke Prediction (https://www.kaggle.com/fedesoriano/stroke-prediction-dataset/code)
• Notebooks: subpopulation_attack_credit_risk.ipynb & Subpopulation_Stroke.ipynb