FHPLF model cuts communication costs 10x with binary gradients in federated learning
Binary matrices replace real-valued gradients to slash costs and boost privacy by 50%
Traditional hash learning requires centralizing user data, conflicting with privacy regulations. Federated learning avoids centralization but transmits large real-valued gradients, causing high communication overhead and privacy risks. Jialan He's FHPLF model addresses both issues by replacing real-valued gradients with binary gradient-like matrices. This drastically cuts computation, storage, and bandwidth costs while providing stronger privacy protection. FHPLF further enhances representation quality with Projected Hamming Distance, which weighs individual binary bits for better similarity modeling, and the Secure Binary Gradient Reassembly and Privacy-Enhanced Upload (SBG-PEU) strategy to minimize user interaction leakage during transmission.
Evaluated on four real-world datasets, FHPLF consistently beats existing hash learning and federated learning methods, achieving a favorable accuracy-efficiency-privacy trade-off. The binary gradient approach reduces communication and storage costs by an order of magnitude while maintaining or improving model accuracy. This makes FHPLF particularly valuable for resource-constrained and privacy-sensitive applications like healthcare, finance, and edge AI, where both data security and bandwidth are critical constraints.
- Replaces real-valued gradient matrices with binary gradient-like matrices, slashing communication and storage costs
- Projected Hamming Distance captures individual bit importance to boost binary code representation capability
- Secure Binary Gradient Reassembly strategy prevents user interaction leakage during transmission
Why It Matters
Enables privacy-preserving federated learning with lower costs, ideal for regulated industries like healthcare and finance.