What is fbgm?

FBGM: Federated Byzantine Gradient Masking

FBGM, or Federated Byzantine Gradient Masking, is a robust and secure Federated Learning (FL) algorithm designed to mitigate the impact of Byzantine attacks during the training process. Byzantine attacks, in the context of FL, refer to malicious or faulty clients that send intentionally corrupted or misleading updates to the central server. These attacks can severely degrade the performance and reliability of the global model trained via FL.

Key Concepts and Mechanisms:

• Federated Learning (FL): FBGM operates within the framework of FL, where a global model is trained collaboratively by a multitude of clients without directly sharing their local datasets. Clients train a model locally on their data and send model updates (e.g., gradients) to a central server. The server aggregates these updates to improve the global model. More information can be found at Federated Learning.

• Byzantine Attacks: FBGM specifically addresses the threat of Byzantine attacks, where malicious clients intentionally send corrupted updates to the server. These updates can be arbitrary and aimed at poisoning the global model, making it perform poorly or even behave maliciously. Understand Byzantine Attacks in detail.

• Gradient Masking: The core of FBGM lies in the gradient masking technique. Each client applies a mask to their local gradient before sending it to the server. This mask is designed to obscure the gradient information, making it difficult for Byzantine attackers to inject malicious updates without being detected. More about Gradient Masking.

• Server-Side Aggregation: The server receives masked gradients from all participating clients. It then performs a robust aggregation method, such as the median or trimmed mean, to mitigate the influence of potentially malicious updates. These robust aggregation techniques are less sensitive to outliers or corrupted values. Details are available about Server-Side Aggregation.

Benefits of FBGM:

• Robustness to Byzantine Attacks: FBGM is designed to be resilient against a significant number of Byzantine attackers, ensuring the integrity and accuracy of the trained global model.

• Privacy-Preserving: Like other FL algorithms, FBGM protects client privacy by avoiding direct sharing of local data.

• Improved Model Accuracy: By mitigating the impact of Byzantine attacks, FBGM helps maintain the accuracy and performance of the global model.

Limitations:

• Computational Overhead: Applying gradient masking and using robust aggregation methods can introduce some computational overhead.

• Communication Costs: The communication costs might increase due to the need to transmit masked gradients.