From Adversarial Attacks to Robust Classifiers- A Study in Social Media Spam Detection

The rise of social media platforms has transformed communication, but also introduced challenges such as spam and phishing. Of particular concern are adversarial attacks, which subtly modify data in an effort to subvert machine learning-based spam detection systems. This study addresses these challenges by building a novel dataset of social media messages reflecting the most common adversarial attacks, including synonym replacement, character-level perturbations, and spammy word insertion techniques. Seven conventional machine learning models, including Logistic Regression, Random Forest, and Gradient Boosting, were evaluated for their performance in detecting spam under adversarial conditions. Logistic Regression achieved the highest accuracy of 99%, demonstrating robustness and reliability in spam detection. This work underscores the critical importance of designing resilient spam detection systems and provides a benchmark dataset and methodology for future research in adversarial machine learning.