MUMBAI, India, June 22 -- Intellectual Property India has published a patent application (202621048323 A) filed by Dr. D. Y. Patil Institute Of Technology, Pimpri, Pune - on April 16, 2026, for Compression-Friendly Secure Aggregation In Federated Learning.
Inventors include Ms. Sanskruti Marathe; Mr. Sharad Adsure; and Ms. Shraddha Shingne.
The application for the patent was published on June 12, 2026, under issue no. 24/2026.
Abstract: Federated Learning (FL) allows multiple parties to train machine learning models on data stored locally, which helps address privacy concerns. However, two major challenges complicate its use: high communication costs due to large model updates and privacy risks from possible data leaks. Most research looks at these problems separately. Communication efficiency improves through compression methods like quantization and sparsification. Meanwhile, privacy concerns are dealt with by secure aggregation methods that use cryptography to protect individual contributions. This creates a fundamental conflict because these solutions do not work well together. Secure aggregation methods usually need dense, fixed-size vectors and do not handle client dropouts well, which often happen in real-world situations. Compression methods, on the other hand, typically produce sparse or variable-length updates that are not compatible with these secure protocols. This research introduces a new secure aggregation framework that supports compression while being efficient, private, and robust. Our hybrid approach combines three main parts. First, clients compress their local updates aggressively to greatly lower communication costs. Second, we use a dropout-resilient cryptographic masking method on these smaller, compressed updates to ensure privacy, even when a significant number of clients (up to 30%) disconnect. Lastly, to ensure the model maintains high accuracy and converges properly, we implement an error-feedback mechanism that carries the compression error into the next rounds. By addressing these issues together, this work offers a practical and scalable way to apply FL in areas with limited resources and strict privacy needs, like healthcare, finance, and the Internet of Things (IoT). Keywords: Federated Learning, Secure Aggregation, Model Compression, Communication Efficiency, Privacy Preservation, Dropout Resilience, Error Feedback, Convergence Guarantees.
Disclaimer: Curated by HT Syndication.
