Large-scale data collection is challenging in alternative centralized learning as privacy concerns or prohibitive policies may rise. As a solution, Federated Learning (FL) is proposed wherein data owners, called participants, can train a common model collaboratively while their privacy is preserved. However, recent attacks, namely Membership Inference Attacks (MIA) or Poisoning Attacks (PA), can threaten the privacy and performance in FL systems. This paper develops an innovative Adversarial-Resilient Privacy-preserving Scheme (ARPS) for FL to cope with preceding threats using differential privacy andcryptography. Our experiments display that ARPS can establish a private model with high accuracy out performing state-of-the-art approaches. To the best of our knowledge, this work is the only scheme providing privacy protection beyond any output models in conjunction with Byzantine resiliency without sacrificing accuracy and efficiency.

Federated Learning enables aggregating models trained over a large number of clients by sending these models to a central server, while data privacy is preserved since only the models are sent. Federated learning techniques are considerably vulnerable to poisoning attacks. In this paper, we explore the threat of poisoning attacks and introduce a game-based robust federated averaging algorithm to detect and discard bad updates provided by the clients. We model the aggregating process with a mixed-strategy game that is played between the server and each client. The valid actions of the clients are to send good or bad updates while the server can accept or ignore these updates as its valid actions. By employing the Nash Equilibrium property, the server determines the probability of providing good updates by each client. The experimental results show that our proposed game-based aggregation algorithm is significantly more robust to faulty and noisy clients in comparison with the most recently presented methods. According to these results, our algorithm converges after a maximum of 30 iterations and can detect 100% of the bad clients for all the investigated scenarios. In addition, the accuracy of the proposed algorithm is at least 15.8% and 2.3% better than state of the art for flipping and noisy scenarios, respectively.

Credit risk is one of the major challenges faced by all financial institutions. Different institutions apply various techniques and models to reduce the risks associated with lending and other financial activities. However, due to the sensitivity of financial data and the diversity of modeling approaches, sharing data among institutions is extremely difficult, often impossible. As a result, improvements in credit risk prediction models typically occur in isolation, hindering collective progress toward higher accuracy and broader effectiveness. Federated learning offers a promising solution by allowing institutions to collaboratively train models without exposing or transferring sensitive data. In this research, we present a federated learning architecture for credit risk prediction that ensures privacy throughout the entire training process. Our results indicate that this approach not only protects data confidentiality but also maintains high predictive accuracy over numerous training rounds, offering a reliable and efficient framework for institutional adoption. The core contribution of this work is the development of a decentralized federated learning (FL) architecture tailored to heterogeneous, non-IID financial data. This framework enhances privacy, scalability, and regulatory compliance, and demonstrates performance advantages over traditional methods. In this article, we demonstrate that using five real-world credit risk datasets, the decentralized FL architecture significantly improves model accuracy (ranging from 71% to 99%) compared to traditional machine learning methods, especially in scenarios where privacy and communication efficiency are essential. While centralized FL achieves the highest average accuracy (up to 83%), the decentralized model provides a strong trade-off between performance and privacy-aware collaboration.

With the modern invention of high-quality sensors and smart chips with high computational power, smart devices like smartphones and smart wearable devices are becoming primary computing sources for routine life. These devices, collectively, might possess an enormous amount of valuable data but due to privacy concerns and privacy laws like General Data Protection Regulation (GDPR), this enormous amount of very valuable data is not available to train models for more accurate and efficient AI applications. Federated Learning (FL) has emerged as a very prominent collaborative learning technique to learn from such decentralized private data while reasonably satisfying the privacy constraints. To learn from such decentralized and massively distributed data, federated learning needs to overcome some unique challenges like system heterogeneity, statistical heterogeneity, communication, model heterogeneity, privacy, and security. In this article, to begin with, we explain some fundamentals of federated learning along with the definition and applications of FL. Subsequently, we further explain the unique challenges of FL while critically covering recently proposed approaches to handle them. Furthermore, this paper also discusses some relatively novel challenges for federated learning. To conclude, we discuss some future research directions in the domain of federated learning.

A decentralized method of machine learning, federated learning (FL) enables several clients to work together to train models without disclosing their raw data. However, because of its openness, it is also susceptible to poisoning assaults, especially label-flipping (LF), in which harmful clients alter training labels to taint the global model. Such effort drifts the model in a way that the model performance dwindles in specific attack-related classes while behaving the same as benign clients for other classes to increase complexity for detecting solutions. We combat this by using a defense mechanism that dynamically modifies trust factors to filter out malicious updates based on last-layer gradient similarity. By assessing the defense across a variety of datasets and more complex adversarial scenarios, such as multi-group attacks with different intensities, this study builds on earlier studies.  According to experimental data, the method maintains accuracy within a proper level of the clean model while drastically reducing the impact of label-flipping, cutting the attack success rate by 50%.  These results demonstrate how important it is to have adaptive security measures in place to protect FL models in hostile and changing contexts.