Early and accurate detection of Autism Spectrum Disorder (ASD) is crucial for optimal intervention and support. Traditional diagnostic methods often face challenges such as social stigma, limited access to specialized services, and difficulties in assessing nonverbal children. This research explores the potential of Federated Learning (FL) to address these limitations and improve ASD diagnosis. By training models collaboratively on decentralized data without compromising patient privacy, FL offers a promising approach to enhance early detection. Our study developed and evaluated a novel FL framework, incorporating three deep learning models to analyze various ASD-related features. Results demonstrate the feasibility and effectiveness of the proposed method, achieving a peak accuracy of 97.5%, surpassing previous studies. This research highlights FL’s potential to revolutionize ASD diagnosis, enabling more accessible, inclusive, and accurate assessments while safeguarding sensitive patient information. By overcoming the challenges of traditional methods, FL can contribute to earlier interventions and improved outcomes for individuals with ASD.