Enhancing Support Vector Machines Against Poisoning Attacks: A Study on Noise Augmentation for Robust Modulation Detection

The proliferation of machine learning applications has led to an increased risk of poisoning attacks, where adversaries inject harmful data into training sets to degrade model performance. Support Vector Machines (SVMs) are particularly vulnerable due to their sensitivity to outliers. This study investigates the application of noise augmentation as a defense strategy against poisoning attacks on SVMs within the context of automatic modulation detection in communication systems. Using the RADIOML2016.10a (RML2016.10a) dataset, we evaluate the impact of noise augmentation on the robustness of SVM classifiers under adversarial conditions. Results indicate that introducing noise points into the training data significantly improves the model's accuracy and Probability of Detection (POD) at low False Positive Rates (FPR), thereby enhancing resilience to adversarial manipulations. The study shows a 50% and 40% improvement in accuracy post-data poisoning attack with the addition of white Gaussian noise and uniform random noise to the training data respectively. These findings show the potential of noise augmentation to improve the resilience of SVMs against data poisoning attack, contributing to more secure and reliable communication systems.