Securing ML Models: A Systematic Survey of Poisoning Attacks and Defense Mechanisms

In recent years, Machine Learning (ML) has brought about a significant revolution in several fields such as medicine, justice, cybersecurity, and other vital fields that require intelligent and urgent decision-making. With this development, a type of adversarial attack targeting ML models called a Poisoning Attack (PA) has emerged. One realistic attack scenario is for an adversary to subtly update samples or reverse some labels of training data, causing degradation to the model's overall accuracy during the testing phase. To gain a deeper understanding of this scenario, a survey will be conducted about the attack and how it is carried out against different models. In addition to the protection techniques to identify their weaknesses. Finally, some solutions will be proposed to maintain the availability, robustness, and integrity of ML models.