Machine Learning is a part of every industry accross the worlds and thereby also imapcts our everyday lives. Though most of the times the beneficial side of these technologies are shown, yet some adversaries rises. Once such adversary is the leakage of sensitive data In the python notebook : CIS545_Fall22_ModelConfidence.ipynb

CIFAR 10 dataset has been used to establish a mean of model securiry. The aim was to divide the member data into 5 disjoint set and run them through a base model and derive their prediction confidence. The model which resulted in a higher prediction accuracy was then eliminated from model training to ensure that the attack result's accuracy would be dropped.

The afore mentioned notebook can be ran sequentially and has been sectioned as below:

Relavent libraries installations: Library Imports Load Dataset - CIFAR 10 Data Explorations & Pre-processing Machine Learning Model Generating Subset of the Dataset Model Tuning Generate Prediction Vector ATTACK 1 (Threshold attack) on each model ATTACK 2 (MLP attack) on each model Plotting of attack result to visualize the accuracy

Attack on a Combination of all model probability vectors [with highest confidence]

Exclusion Oracle generation Attack on the oracle and attack accuracy visualization

The additional notebook : CIS545_Fall22_ModelConfidence_increased_epoch.ipynb is for the purpose of displaying how increasing the iterations while training the model helps in improving privacy.