Defending the Integrity of Machine Learning Models in the Face of Adversarial Threats

Abstract

Abstract The primary objective of the research is to examine the resilience of models based on machine learning when confronted with adversarial threats. The focus will be on developing effective strategies to safeguard the integrity of these models. This study employs a carefully designed framework to undertake the task of simulating adversarial attacks on machine learning models. The central objective of the research centres on the intentional manipulation of numerical variables in order to accurately simulate real-world threats. The investigation of statistical analyses, such as t-tests and comprehensive data visualisations, enables the evaluation of the impact exerted by adversarial attacks on both data distributions and model outputs. The primary objective of the current investigation is to assess the effectiveness of a mimicked defence mechanism in the domain of machine learning model restoration, with the ultimate goal of preserving model integrity. The results gathered by our study emphasise the significant significance of adversarial threats, thereby underscoring the need for a comprehensive analysis of robust defence strategies. The analysis of correlation matrices, both before and after adversarial attacks, provides valuable insights into the changes in the fundamental structure induced by these interventions. The current investigation offers valuable insights into the preservation of the integrity of machine learning models and suggests possible ways to enhance their resilience against ever-changing adversarial tactics.