| Deep learning has been widely adopted for developing intrusion detection systems in smart grid networks. Privacy of data generated by smart meters and other grid networks is a concern for such data driven techniques. Federated learning is a framework of distributed learning in these scenarios, where embedding or model parameters are shared across edge devices instead of sensitive data points. Unknown zero-day attacks are common in such a distributed smart grid. Zero-shot learning provides a way of handling such attacks. In the context of intrusion detection a textual description of the attacks provides auxiliary information helpful in recognizing zero-day attacks. Large language models are often useful in generating this auxiliary information. We propose a federated zero-shot learning framework where a global LLM generates privacy insensitive auxiliary text information about unknown attacks, and the corresponding embedding is shared across federated learning clients. The clients use deep neural networks for attack detection. The approach can detect zero-day attacks while maintaining the privacy of consumer-sensitive data. We conduct experimental studies on the benchmark Electra intrusion dataset. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.