Abstract
Traditional network security analysis methods exhibit critical limitations in processing high-dimensional dynamic data, including inefficient feature selection, poor adaptability to evolving threats, and low detection sensitivity below 50%. To address these challenges, this study proposes a multi-objective multi-label feature selection model integrated with an optimized Fireworks Algorithm. The Improved Fireworks Algorithm Model incorporates Gaussian operators and adaptive functions while fusing fuzzy neural networks to enhance real-time threat response. Experimental validation across Palmer Penguin (small-scale), Fashion MNIST (medium-scale), and Bike Sharing (large-scale) datasets demonstrates three key advancements: Data processing capacity reaches 5,000 samples, exceeding Particle Swarm Optimization and standard Fireworks Algorithm baselines by 66%; Sensitivity maintains 70%-100% across datasets, outperforming traditional methods by 30% points; In a medium-sized data set, the research method scored only 5 out of 10 in the five indicators of comprehensive performance comparison based on the weighted geometric mean of the five-dimensional radar chart, indicating that the research method may have problems of overfitting or insufficient generalization ability when processing complex data. Adaptive adjustment time is reduced by 50%, confirming significant efficiency gains. These findings establish a robust framework for dynamic network security while highlighting scalability constraints in complex data environments.