A Critical Study on the Application of Explainable AI for Handling Data Poisoning Attacks

Abstract

With the rapid integration of machine
learning models into various critical
applications, the susceptibility of these models
to adversarial attacks has emerged as a
significant concern. Data poisoning attacks, a
subset of adversarial attacks, involve injecting
malicious or misleading data into the training
set with the intent to degrade the model's
performance or induce erroneous predictions.
Explainable Artificial Intelligence (XAI)
techniques have gained prominence as a
promising approach to enhance the resilience
of machine learning models against such
attacks. This paper explores the role of
Explainable AI in mitigating data poisoning
attacks by providing interpretable insights
into the model's decision-making process. We
review various XAI methods that aid in
detecting and mitigating the presence of
poisoned data during both the training and
inference phases. Additionally, we discuss how
XAI techniques facilitate the identification of
attack vectors, aiding in the development of
more robust models. Through experimental
evaluations and case studies, we demonstrate
the efficacy of XAI in enhancing model
security and reliability. The findings
underscore the importance of integrating
Explainable AI strategies into the model
development pipeline to bolster defence
mechanisms against data poisoning attacks,
thereby fostering trust and dependability in
AI systems across diverse applications