Image-to-image translation has become a prominent trend in the field of computer vision. This innovative technique is widely employed for generating concealed facial features from given noise. It proves particularly useful in reconstructing obscured facial regions covered by masks. The underlying technology at the core of this method is the Generative Adversarial Network (GAN). The primary mechanism of GAN-based neural networks relies on an encoder-decoder structure combined with a discriminator network. The discriminator plays a crucial role in training the network to distinguish between authentic and generated images. FUGAN, a novel approach in this domain, incorporates a self-attention block and has demonstrated superior performance to other models. With an impressive SSIM and BRISQUE score of 0.928 and 27.990, respectively, FUGAN is at the forefront of achieving state-of-the-art results. The architecture of the FUGAN neural network is built upon the UNET framework, with modifications to enhance its effectiveness. These modifications involve increasing the network depth and incorporating batch instance normalization as a regularizer to achieve the desired outcomes. The resulting model attains a smaller size and exhibits high accuracy in image regeneration. This streamlined and efficient FUGAN model is well-suited for deploying low-resource Internet of Things (IoT) devices. Its capability to regenerate facial features from masked regions makes it instrumental in surveillance systems, detecting individuals wearing masks and reconstructing their facial details. Github link (FUGAN: https://github.com/mrinmoy-sadhukhan/FUGAN-model ) of our work is provided.