Breaking classification Algorithms: A Black-Box Adversarial Attack for Data Poisoning

Poisoning attacks interfere with machine learning classifiers' performance by introducing malicious data into the learning dataset. In this paper, we propose a black-box adversarial attack for data poisoning in machine learning, where the attacker has no knowledge of the classifier's internal parameters or training data. Our attack utilizes a novel poison threat algorithm that leverages the gradient descent method to generate adversarial examples that deceive the classifier. We evaluate our attack on two popular classification algorithms, decision trees and support vector machines, using the random circles dataset. According to our results, the proposed attack has the ability to compromise the performance of both classifiers, highlighting the need for robust countermeasures against poisoning attacks.