Memes, with their fusion of images and text, have become a cornerstone of digital communication, encapsulating humor, cultural critique, and social commentary. However, their potential to disseminate offensive or harmful content presents a formidable challenge for automated content moderation systems, which often struggle to decipher the complex interplay between visual and textual elements. This study proposes an innovative multimodal deep learning framework to identify offensive memes, utilizing a robust dataset of annotated memes designed to test the synergy of text and image modalities. The approach employs the Inception-ResNet-V2 model, an advanced convolutional neural network, to extract intricate visual features from meme images, complemented by a transformer-based model that captures nuanced textual semantics. These modalities are integrated through a late-fusion strategy, enabling the model to interpret combined meanings that elude unimodal systems. Experimental evaluation reveals a balanced performance, achieving an overall accuracy of 55% and a macro-averaged F1-score of 0.51. The framework demonstrates notable strength in detecting non-offensive content, with a recall of 0.83, indicating reliability in identifying benign memes. However, its lower recall of 0.26 for offensive content highlights the difficulty of capturing subtle or context-dependent harmful intent. These findings illuminate the intricacies of multimodal classification and underscore the need for advanced techniques to address semantic ambiguities. By enhancing the detection of offensive content, this research contributes to the development of more effective content moderation tools, fostering safer and more inclusive online environments. It also lays a foundation for future explorations into real-time applications and cross-cultural adaptations, addressing the evolving landscape of digital communication.