Intelligent Attacks and Defense Methods in Federated Learning-enabled Energy-Efficient Wireless Networks

  • 2025-04-25 18:40:35
  • Han Zhang, Hao Zhou, Medhat Elsayed, Majid Bavand, Raimundas Gaigalas, Yigit Ozcan, Melike Erol-Kantarci
  • 0

Abstract

Federated learning (FL) is a promising technique for learning-based functionsin wireless networks, thanks to its distributed implementation capability. Onthe other hand, distributed learning may increase the risk of exposure tomalicious attacks where attacks on a local model may spread to other models byparameter exchange. Meanwhile, such attacks can be hard to detect due to thedynamic wireless environment, especially considering local models can beheterogeneous with non-independent and identically distributed (non-IID) data.Therefore, it is critical to evaluate the effect of malicious attacks anddevelop advanced defense techniques for FL-enabled wireless networks. In thiswork, we introduce a federated deep reinforcement learning-based cell sleepcontrol scenario that enhances the energy efficiency of the network. We proposemultiple intelligent attacks targeting the learning-based approach and wepropose defense methods to mitigate such attacks. In particular, we havedesigned two attack models, generative adversarial network (GAN)-enhanced modelpoisoning attack and regularization-based model poisoning attack. As acounteraction, we have proposed two defense schemes, autoencoder-based defense,and knowledge distillation (KD)-enabled defense. The autoencoder-based defensemethod leverages an autoencoder to identify the malicious participants and onlyaggregate the parameters of benign local models during the global aggregation,while KD-based defense protects the model from attacks by controlling theknowledge transferred between the global model and local models.

 

Quick Read (beta)

loading the full paper ...