Resilience of Federated Learning Against False Data Injection Attacks in Energy Forecasting

Federated learning (FL) has established itself as a communication-efficient, privacy-aware, and cost-effective technique for training machine learning models in energy forecasting. This approach enables simultaneous model training across multiple smart grids while keeping data decentralized at edge nodes. However, FL is not immune to backdoor adversarial attacks, such as data and model poisoning. In this paper, we scrutinize the impact of two data poisoning techniques: scaling and random noise effects. The attack was initiated on one client among ten. As the attack percentage increases, the Mean Absolute Percentage Error (MAPE) of the local model also rises. Our simulation results reveal that the scaling effect elevated MAPE from 0.193% to 32.72%, while random noise increased MAPE from 0.183% to 129.75% as the attacked percentage rose from 10% to 100%. It is concluded that data poisoning solely affects the local model and does not significantly impact the global model; hence, it can provide more resilience than centralized machine learning models.