IoT networks are increasingly vulnerable to adversarial attacks, necessitating robust detection mechanisms to ensure security. This study proposes a comprehensive framework for identifying adversarial attacks in IoT systems using ML and DL models. Leveraging the RT-IOT2022 dataset, we preprocess network traffic data through class balancing, feature selection, and label encoding to mitigate imbalances and enhance model performance. Four classification models—CNN, GRU, Random Forest, and Stacking Classifier—are developed and evaluated for multi-class attack detection. The Stacking Classifier achieves the highest accuracy of 96%, followed by Random Forest at 95% and GRU at 92%, while CNN yields 66% due to feature scaling limitations. A Flask-based web application is deployed to enable real-time attack prediction, integrating the Stacking Classifier with a user-friendly interface and MySQL database for user management. This work demonstrates the efficacy of hybrid ML approaches in securing IoT networks and highlights the need for improved feature preprocessing for DL models. Future enhancements include adversarial training and real-time data integration to strengthen robustness against evolving threats.