Impact of Data Poisoning Attack on the Performance of Machine Learning Models

The twenty-first century has witnessed widespread adoption of Artificial Intelligence (AI), Machine Learning (ML) and Deep Learning (DL). These techniques have provided reliable solutions in various areas, including statistics, information theory, and mathematics. Given the prevalence of ML techniques, there exist various adversaries which question the robustness of ML models. Adversaries aim to manipulate models to their advantage, reducing their performance and accuracy. Data poisoning attack is one such adversary in which the attacker manipulates models by introducing specially crafted poisoned data into the training dataset. This paper presents the performance analysis of different machine learning models with and without the influence of data poisoning attack to predict the probability of diabetes and its effect on accuracy and precision. It has been observed that the SVM (RBF) classifier performs best on clean data, while the KNN classifier is highly affected by data poisoning, with a lesser impact on the SVM (Linear kernel) classifier.