The need for strong defenses against poisoned attacks has been highlighted by the growing application of machine learning in vari- ous fields. This attack alters training data to cause the model to be mis- classified. This study examines and compares several defense strategies, such as TRIM, RANSAC, and N-LID, to reduce the impact of poisoning attacks. Large-scale poisoning attacks were shown to be best addressed by TRIM, which excludes outliers during training using a trimmed loss function. Furthermore, N-LID was used to detect subtle attacks by calcu- lating the local intrinsic dimensionality of each data point and applying a weighting mechanism to reduce the impact of vulnerable points. The study shows that the hybridization of TRIM and N-LID offers the best defense solution, outperforming other defenses in terms of model accu- racy and robustness. This approach ensures that machine learning models remain resilient against both large-scale and subtle adversarial threats while maintaining their performance on non-poisoned data.