Detecting Poisoning Attacks with DynaDetect

In the age of increasing reliance on machine learning (ML) in various environments, ensuring the security and reliability of ML models is essential. Data poisoning attacks pose a significant threat to ML models, compromising their reliability. Although the traditional K-Nearest Neighbor (KNN) algorithm can offer potential defense mechanisms due to its adaptability and capability to detect poisoned data, its static nature limits its effectiveness against dynamic malicious challenges. To this end, this work proposes DynaDetect, a dynamic KNN-based algorithm designed to detect data poisoning attacks. Our methodology adapts the traditional KNN model to a dynamic framework, allowing to adjust its parameters, such as the number of neighbors considered, based on the characteristics of the data. The experimental results indicate a marked improvement in the detection accuracy of poisoned data, enhancing the reliability of ML models.