# Dynamic Defense Against Byzantine Poisoning Attacks in Federated Learning

**Authors:** Nuria Rodríguez-Barroso, Eugenio Martínez-Cámara, M. Victoria Luzón, Francisco Herrera  
**arXiv:** [2007.15030](https://arxiv.org/abs/2007.15030)  
**Date:** 2020-07-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.15030v2/

## Key points

- Vulnerability of federated learning models to Byzantine poisoning attacks
- Proposal of a novel approach called Dynamic Defense Against Byzantine Poisoning Attacks (DDaBA)
- DDaBA dynamically filters out adversarial clients during the aggregation process in real-time
- Effectiveness of DDaBA demonstrated through experiments on image datasets like Fed-EMNIST Digits, Fashion MNIST, and CIFAR-10
- Improvement in global learning model performance by dynamically selecting clients for aggregation
- Enhanced security and robustness of federated learning systems against malicious attacks

## Summary

In their paper titled "Dynamic Defense Against Byzantine Poisoning Attacks in Federated Learning," Nuria Rodríguez-Barroso, Eugenio Martínez-Cámara, M. Victoria Luzón, and Francisco Herrera address the vulnerability of federated learning models to . These attacks can compromise the integrity of the global learning model by corrupting it with malicious data from . To mitigate this threat, the authors propose a novel approach called . The key idea behind DDaBA is to dynamically filter out adversarial clients during the aggregation process in federated learning. Traditional federated aggregation operators have been ineffective at preventing such attacks or rely on assumptions about the nature of adversarial clients. In contrast, DDaBA dynamically adjusts its aggregation strategy to discard potentially malicious clients in real-time. To evaluate the effectiveness of DDaBA, the authors conducted experiments using deep learning classification models on popular image datasets like Fed-EMNIST Digits, Fashion MNIST, and CIFAR-10. The results demonstrate that dynamically selecting clients for aggregation significantly improves the performance of the global learning model while effectively excluding adversarial and low-quality clients. Overall, this dynamic defense mechanism offers a promising solution to enhance the security and robustness of federated learning systems against . The findings presented in this study contribute valuable insights towards developing more  frameworks in real-world applications.