Dynamic Gradient Compression and Attack Defense Strategy for Privacy Enhancement of Heterogeneous Data in Federated Learning

An innovatively constructed dynamic perception mechanism based on temporal and spatial dual dimensions is proposed. In particular, it dynamically adjusts the gradient compression ratio depending on the convergence rate of the model (e.g., using high compression ratio for fast convergence at early stage and fine-tuning later), which is realized by integrating loss variation rate and training cycle into compression ratio equation. In spatial dimension, this method adapts to heterogeneous data distributions by introducing a sample weight factor into the non-IID measurement index so that the heterogeneity of data can be quantified. A dynamic privacy budget allocation strategy based on data sensitivity matrix ensures adaptive noise injection and hierarchical encryption. In contrast to traditional methods, the anomaly detection module introduces high order statistical moments (skewness, kurtosis), combined with machine learning based attack classification methods, to detect gradient poisoning and model stealing attacks in real time.