WebGPU Accelerated Client-Side AI for Privacy Preserving Dermatological Diagnostics: Performance Benchmarking and Local Differential Privacy Integration
Abstract: Clinical deep learning traditionally relies on server-centric architectures, raising critical concerns regarding patient data privacy and latency. This research presents an empirical benchmarking study of on-device skin lesion classification using the WebGPU API. By utilizing native compute shaders for tensor operations, I evaluate the inference performance of MobileNetV2 and EfficientNetV2 architectures directly within the web browser. My benchmarks demonstrate that WebGPU provides a 3.7x speedup for MobileNetV2 image classification tasks compared to WebGL 2.0 1, while maintaining high diagnostic accuracy. To safeguard patient confidentiality, I integrate a Local Differential Privacy (LDP) layer with a privacy budget of ?=1.9, which results in a marginal utility trade-off consistent with benchmarks in other medical imaging modalities.2 Results indicate that browser-based serverless edge computing can achieve real-time performance (60 FPS) on mobile devices without the thermal throttling associated with legacy APIs. This study establishes a viable path for scalable, privacy-preserving dermatological screening on consumer-grade hardware.
Keywords: Client-side AI, WebGPU, Privacy Preserving Machine Learning, Deep Learning, Medical Imaging
How to Cite?: Arpankumar Patel, "WebGPU Accelerated Client-Side AI for Privacy Preserving Dermatological Diagnostics: Performance Benchmarking and Local Differential Privacy Integration", Volume 15 Issue 2, February 2026, International Journal of Science and Research (IJSR), Pages: 1191-1193, https://www.ijsr.net/getabstract.php?paperid=SR26219113252, DOI: https://dx.doi.org/10.21275/SR26219113252