Random Data Poisoning Attacks on Nonlinear Regression Learning

Nonlinear regression has numerous applications in diverse fields, including biology, economics, engineering, and more, where it is used to model and analyze complex relationships between variables that cannot be adequately represented by linear models. However, these models are susceptible to malicious attacks that manipulate input data to yield false results. This study focuses on unpredictable data poisoning threats based on randomization in nonlinear regression learning and assesses the iTrim defense mechanism’s efficacy. Multiple nonlinear regression datasets and common techniques were used in experiments. Random Data poisoning attack involves regenerating data points with altered labels and inserting them into the training set. The polluted dataset underwent iTrim defense, and model performance on a test set gauged effectiveness. Results show that models suffer significant performance degradation when exposed to random data poisoning attacks. Malicious points cause overfitting and poor test set generalization. This study underscores nonlinear regression models’ vulnerability to random data poisoning and the need for robust security measures, while iTrim offers some protection, further research is vital to develop more potent defense systems against complex attacks.