On August 14, the "Symposium on Cloud-Edge Collaborative Computing in the Era of LLMs" was held at Fudan University, Shanghai. The event was co-organized by the School of Computing and Intelligence Innovation at Fudan University and CISTER Research Center, Portugal. More than twenty scholars and industry experts from China, Portugal, Germany, the United Kingdom, and other countries gathered to engage in in-depth discussions on the new architectures, practical implementations, and security challenges of cloud-edge collaborative computing in the context of large language models (LLMs).
One of the symposium's co-organizers, Prof. Jin Zhao of Fudan University, emphasized that cloud-edge-user collaborative computing has become a critical pathway for enhancing computational efficiency and ensuring training and inference performance in the era of LLMs. He introduced his team’s newly developed framework, FedFreeze, designed for federated learning training. Tailored for fine-tuning in resource-constrained scenarios, FedFreeze achieves dual optimization of accelerated convergence and reduced memory consumption through globally coordinated selective tensor freezing, fine-grained gradient operations, and gradient compression. The team has validated the framework's superior performance across various heterogeneous hardware platforms and diverse datasets.
Dr. Kai Li from CISTER Research Center presented a data-agnostic paradigm for model poisoning in federated learning. Using an adversarial graph autoencoder (GAE), the method disrupts federated training without accessing the original training data. Moreover, he proposed a defense mechanism based on explainable graph neural networks, which integrates Grad-CAM and autoencoders to achieve robust detection of GAE-based poisoning attacks. This approach offers stronger interpretability and resilience compared to traditional Euclidean distance-based methods.
In the concluding remarks, Prof. Xiaoming Fu, a member of the German National Academy of Science and Engineering and the European Academy of Sciences, a professor at the Institute of Computer Science at the University of Göttingen, and Chief Scientist at the Social Intelligence Research Center of Fudan University, noted that the symposium facilitated in-depth exchanges on cutting-edge topics, such as intelligent wireless network environments, cloud-edge-user collaborative scheduling, resilience and efficient fine-tuning in federated learning, LLMs inference via cloud-edge collaboration, video intelligence analysis, and observability in ultra-large-scale GPU cluster LLM training. These discussions are expected to promote mutual empowerment between LLMs and cloud-edge collaborative computing. He highlighted that future advancements in cloud-edge collaborative computing will continue to break new ground in intelligent computational scheduling, cross-domain resource integration, and data privacy protection. The complementary strengths of China and Europe in standards, technology, and applications are likely to inject new momentum into the development of a global intelligent computing system through further collaboration.
