context and background

Federated Learning enables decentralized training of machine learning models on data that remains with clients, ensuring privacy and data security

Overall problems framework with federated learning

significant challenges in Federated Learning (FL), particularly focusing on balancing privacy, transparency, and model explainability within AI systems. The key problems identified include:

1. Privacy Requirements Limiting Data Access: FL was designed to address privacy concerns by allowing multiple data owners to collaboratively train models without sharing raw data. However, this decentralization restricts access to comprehensive datasets, which can reduce model performance. This issue is especially relevant in fields where data sensitivity is paramount, such as healthcare and finance(information-infusion).
2. Explainability and Trustworthiness of Models: A critical requirement for trustworthy AI, particularly in high-stakes applications, is model explainability. Traditional FL models, especially those based on complex neural networks, are often opaque or "black-box" models, lacking transparency. This problem undermines user trust and limits adoption in domains where understanding model decisions is essential(information-infusion).
3. Complexity in Federated Explainable Models: While explainable models like decision trees and fuzzy regression trees offer interpretability, adapting these models to an FL setup introduces challenges, as they don’t rely on simple optimization techniques used in neural networks. Existing methods, such as FedAvg, are not directly applicable, necessitating customized aggregation approaches to maintain interpretability and performance(information-infusion).

These challenges are crucial because they address the core principles of privacy and transparency, which are essential for user trust and regulatory compliance in AI systems.

specific issues and solutions to this will be studied and compared

the security and robustness challenges in Federated Learning (FL), specifically the vulnerability of FL models to backdoor attacks

existing defenses—such as statistical-based, filter-based, and differential privacy approaches—often fall short against sophisticated backdoor attacks, especially under continuous attacks where malicious updates are introduced in every training round

Traditional FL approaches face obstacles such as communication bottlenecks, staleness of updates, and non-IID (non-independent and identically distributed) data among IoT devices.

the challenge of implementing secure and efficient machine learning (ML) models within resource-constrained IoT networks. Given the rise of cyberattacks targeting IoT devices and the challenges posed by encrypted network traffic, existing centralized ML solutions have proven inadequate for IoT environments. Traditional centralized approaches often struggle to analyze and detect malicious behaviors in real-time, especially in distributed and encrypted IoT settings