By integrating Microsoft's latest GraphRAG technology, I successfully built a retrieval-augmented generation system based on knowledge graph construction and the DeepSeek large model. Utilizing the information extraction capabilities of LLMs, I designed prompts to interact with the LLM, significantly improving the quality of answers in the power domain.

An InstructPix2Pix fine-tuning method based on reinforcement learning is proposed to make the Model more strictly follow the editing instructions, reduce unexpected changes and improve the overall editing image quality. We combine the training process of the Diffusion Model with reinforcement learning, and adopt the near-end strategy optimization (PPO) algorithm. A target reward function combining CLIP‑Score and background difference is designed, and a new data set is generated using GPT‑4 prompt words on LVIS data set for training. By comparing with Baseline, our method can reduce the MSE of modified image background by 60%. Through observation, it is found that the coincidence between the generated image and the indication is significantly improved

A medical image privacy protection method was designed to retain clinical information of disease diagnosis while protecting privacy information such as gender and age. Differential privacy protection Viewmaker generative network was used to blur medical images, and the objective function based on Focal Loss was designed. Phased training for gender, age, ethnicity and other privacy information · Compared with SOTA method, while maintaining similar privacy protection performance, Disease recognition achieved 81.5% and 80% F1-score records in retinal fundus image BRSET and chest X-ray CheXpert dataset, respectively

Aiming at the problem that the image is attacked by Adversarial Noise so that the neural network cannot correctly identify the image category, the AR generative model based on VQ‑VAE‑2 is adopted to reconstruct the attacked image. Achieve the purpose of defending against adversarial noise attacks and generating high resolution images · By deploying the model on the linux server side for large-scale training on the Imagenet dataset, the resulting model has higher robustness