Gender recognition becomes a very critical task for security agencies while assessing protest activities. At present, with the advent of GPUs, high computing machines, and Deep Convolution Neural Networks (DCCN), automated gender recognition is possible. In this research work, we explore the performance of various DCNN architectures using transfer learning approaches for gender recognition. We performed a detailed ablation study on different input sizes and on different architectures to see the trade-off between latency and the accuracy of the classification. The performance of models tested against standard dataset WIKI, UTKFace, and Adience. We explored VGG-16 and MobileNetV3 architectures for comparison against accuracy and latency parameters in order to select a model suitable for the embedded device considering their low processing and less storage capacity. Experiments conducted using standard architecture against the standard dataset by changing the resolution and fine-tuning it.