Individuals with special needs most of the time find it harder to identify hazards and dangers as well as circumstances that are socially challenging. Hence, they face the risk of falling victim to abuse and violence. In this paper, the main goal is to help people with special needs to more successfully communicate with others and comprehend their surroundings. Machine learning-based solutions are used to help people with special needs in their communication tasks. The proposed machine learning model contains a convolutional layer, attention layer, and Bidirectional long short-term memory (BiLSTM) layer and achieves 99.00% accuracy performance. We applied federated learning to preserve privacy and to help researchers overcome problems they face when dealing with people with special needs.