Deepfake technology, enabled by advances in artificial intelligence and deep learning, has grown exponentially, offering unprecedented abilities to synthesize hyper-realistic audio-visual content. However, its malicious misuse poses significant challenges for information verification, thereby demanding robust detection methods. This paper introduces a novel deep learning framework designed for the detection of deepfake videos by leveraging audio-visual fusion techniques. Recognizing the intricate anomalies present separately in audio and visual data, our approach integrates both modalities to achieve heightened accuracy in discerning genuine from manipulated content. We utilize a two-stream convolutional neural network (CNN) that separately analyzes the spectral properties of audio and the subtle facial discrepancies in video before fusing the obtained features. A composite feature vector is formed from the fusion, which is then input to a deep neural network (DNN) classifier. The combined strength of these modalities not only addresses the current vulnerabilities in existing detection methods but also anticipates the evolving sophistication of deepfake technologies. Preliminary results suggest an impressive enhancement in detection performance as compared to single-modality methods. Additionally, the flexibility of our framework allows for potential future integrations with other modalities and data types.