Cyber-bullying is a significant issue that negatively affects individuals. With the rise of social networking, it has become an even more serious concern. Leveraging state-of-the-art machine learning techniques, we can achieve high accuracy in detecting abusive or harmful language at the character level, which contributes to cyber-bullying. To improve detection accuracy, this research paper employs a deep learning approach using a seven-layer convolutional neural network (CNN) in combination with the N-gram feature selection method, also referred to as Hybrid-CNN (HCNN). This model uses four-stage architecture to effectively detect cyber-bullying at both the word and character levels, even as attackers try to circumvent detection with diverse, uncategorized abusive language. To address this challenge, the research focuses on detecting synonym character-level cyber-bullying, known as Synon-Char-HCNN. In this paper, the proposed HCNN model is trained using multiple synonyms of words with their respective characters, which helps prevent malicious users from sending cyber-bullying content. Experimental results using a real-world dataset show that the proposed approach outperforms traditional methods, providing significantly better performance.