Real Time Adversarial Attack Prevention Framework Utilizing Distortion and Defense Strategies in Machine Learning and Deep Learning Models

Abstract

The objective of this research is to develop and evaluate a machine newlinelearning-based information security model designed to address the newlinevulnerabilities of various machine learning (ML), deep learning (DL), and newlinePyramid Vision Transformer (PVT) models against adversarial attacks, newlinespecifically focusing on Poison and Evasion Attacks. Despite advancements in newlineadversarial defenses, significant challenges persist, particularly regarding the newlinetransferability of attacks across different models, the limited applicability of newlinecurrent defenses in real-world medical datasets, and the absence of newlinecomprehensive hybrid defense frameworks capable of simultaneously newlineaddressing both poisoning and evasion attacks. To tackle these challenges, this newlineresearch rigorously investigates the effectiveness of a variety of models, newlineincluding Logistic Regression, Random Forest, Support Vector Machine (SVM), newlineGradient Boosting, K-Nearest Neighbors (KNN), Recurrent Neural Networks newline(RNNs), Convolutional Neural Networks (CNNs), and Pyramid Vision newlineTransformers (PVT), utilizing diverse medical datasets such as COVID-19, newlinediabetes, kidney disease, heart disease, breast cancer, retinopathy, and brain newlinediseases. The adversarial attack strategies evaluated encompass poison newlineattacks via Label Flipping, Clean Label Attack, Feature-Space Poison Injection, newlineand Gradient Ascent techniques, while evasion attacks are assessed using newlineInput Modification, FGSM, and BGP-MultiModelFool techniques. Additionally, newlinethe research explores a range of defense mechanisms to mitigate these attacks, newlineincluding Spatial Filtering Techniques, Prediction Confidence, Dynamic Layer- newlineWise Weighting, Adaptive Denoising, Data Augmentation with Perturbations, newlineand Multi-Model Feature Fusion (MMFF). This research emphasizes the newlinenecessity of advanced security measures to safeguard ML, DL, and PVT newlinemodels from adversarial manipulations, thereby ensuring the integrity and newlinereliability of AI systems in critical healthcare applications. newline