Facial biometrics are widely employed in access control systems requiring robust security measures. However, they are particularly susceptible to various attacks, with morphing attacks posing significant risks, especially in high-security contexts such as border control. Morphing attacks involve the creation of a composite facial image by blending the facial features of two or more individuals, resulting in an image that closely resembles the identities of the individuals used in the morphing process. The automatic detection of morphing attacks presents a significant challenge, owing to the existence of diverse morphing generation tools and variable image conditions. In this paper, we proposed a novel differential-based morphing attack detection technique using a time-frequency-based convolutional neural network architecture designed to capture the small variations in the unsigned difference of facial embeddings to detect morphing attacks. Extensive experiments were performed on the newly constructed face morphing dataset using four different types of morphing generation tools. The performance of the proposed method was evaluated against three state-of-the-art D-MAD techniques using two distinct evaluation protocols. The results highlight the exceptional effectiveness of the proposed approach in detecting morphing attacks.