Our research focuses on assessing adversarial attacks against cardiovascular disease prediction models in the area of cyber security, specifically applying adversarial attack techniques to
evaluate vulnerabilities of prediction systems.". We will first be using a convolutional neural network (CNN) algorithm to train the model on a balanced structured (Tabular data) dataset of
cardiovascular data. This model will serve as the basis for evaluating the impact of different adversarial attack methods.
We will subject the trained CNN model to four prominent attack algorithms - Projected Gradient Descent (PGD), Carlini-Wagner (C&W), Brendel-Bethge, and Decision Boundary attacks.
For each attack, we will systematically evaluate the loss in model accuracy compared to the original model. We will start with a low attack strength and then gradually increase it to analyze the impact on accuracy metrics like precision, recall, and F1-score.
This comparative analysis will reveal how different attacks vary in their ability to affect model performance.
In addition, we will explore various defense strategies like adversarial training, defense distillation, and gradient regularization to mitigate these attacks. This will provide insights into
effective countermeasures to strengthen model resilience. We will quantify the extent of accuracy recovery after applying defenses against each type of attack.
