Research on personalized compression algorithm for pre-trained models based on homomorphic entropy increase

This paper investigates the deployment challenges of Vision Transformer (ViT) and Large Language Models (LLMs). Vision Transformer captures global information through multi-head attention mechanisms, but its high computational cost limits its application on mobile devices. Although LLMs have achieved breakthroughs in natural language processing, they also face significant deployment challenges. To address these issues, we propose a hierarchical pruning strategy that distinguishes personalized layers from shared layers through compressed sensing and random sampling, significantly reducing model parameters. Experiments show that the hierarchical mechanism effectively balances pruning and accuracy, providing a new direction for deploying efficient and personalized AI models on mobile devices.