Federated Learning (FL) has become a key approach for training machine learning models without sharing raw data. However, when moving toward decentralized federated learning (DFL)—where there is no central server—new security challenges emerge. In such systems, participants (clients) directly exchange model updates, which makes the network more vulnerable to malicious behavior.
The paper “RepuNet: A Reputation System for Mitigating Malicious Clients in DFL” addresses this issue by introducing a reputation-based mechanism to identify and mitigate harmful clients in the network. In DFL, malicious participants can disrupt the training process in several ways, such as sending corrupted updates, delaying communication, or flooding the system with excessive messages. These attacks can significantly degrade model performance and reliability.
RepuNet proposes a solution where each client is assigned a reputation score based on its past behavior. Instead of relying on heavy infrastructures like blockchain—which can introduce scalability and computational overhead—the system uses a more lightweight and adaptive approach. The key idea is simple: clients that behave reliably gain higher reputation, while suspicious or harmful clients are gradually excluded from the training process. This helps maintain the quality of model aggregation and improves overall system robustness.
The paper evaluates RepuNet under different attack scenarios and shows that incorporating reputation into the training process can effectively reduce the impact of malicious participants while preserving learning performance. Overall, RepuNet highlights an important direction for secure and scalable federated learning systems, especially as decentralized architectures become more common in future distributed AI environments.
RepuNet_ A Reputation System for Mitigating Malicious Clients in DFL