Abstract
With the growing adoption of large language models, artificial intelligence, and other
machine learning technologies, these systems are increasingly being integrated
across various industries. The threats targeting these models have also increased.
One such threat is model poisoning, where malicious actors can poison models by
manipulating the learning data, hampering the accuracy of said models. This research
explores the use of artificial immune networks (AINs) as a robust defence mechanism
against such attacks. Drawing inspiration from biological immune systems, AINs are
employed to detect and mitigate adversarial distortions effectively. An Artificial
Immune Network (AIN) is a bio-inspired computational model derived from immune
network theory, mimicking B-cell interactions, attack suppression, and the processes
of cloning and mutation. By exploring how AINs can be used to prevent model
poisoning as a prevention mechanism, this dissertation contributes to the theoretical
advancement in the field of cybersecurity but also offers practical implications for
developing more secure AI systems. The research done in the paper has shown that
artificial immune networks can detect and neutralising harmful behaviours in machine
learning models and artificial intelligent agents. The AINs can detect the poisoned data
from model poisoning, making the models more resilient and combative against new
threats that can emerge.