Federated AI achieves 99.9% accuracy in IoT cybersecurity and fraud detection

Woxsen University researchers have introduced a significant innovation in privacy-preserving artificial intelligence (AI) for cybersecurity and financial fraud detection. The peer-reviewed study, “Enhancing Privacy in IoT-Enabled Digital Infrastructure: Evaluating Federated Learning for Intrusion and Fraud Detection,” published in Sensors, investigates how the flower framework - a federated learning (FL) environment - can empower machine learning while preserving sensitive data.

The research addresses a pressing challenge in the digital age: how to secure and analyze data generated across billions of IoT devices without compromising user privacy. The team explores the use of federated AI algorithms to improve intrusion detection systems (IDS) and credit card fraud detection across distributed digital infrastructure.

How can privacy be preserved in collaborative AI without compromising performance?

The study investigates how AI models can learn from multiple institutions’ data without exposing individual datasets. Traditional centralized machine learning systems require aggregating data in a single location, creating vulnerability to breaches and unauthorized access. Federated learning, however, enables decentralized model training - clients train AI models locally and share only the learned parameters with a global model.

The flower framework emerged as the preferred infrastructure in the study due to its low communication overhead, enhanced security, and flexibility. Compared to alternatives such as TensorFlow Federated and PySyft, the flower framework outperformed its peers in bandwidth efficiency, convergence speed, and client anonymity. It also maintained strong compatibility with machine learning libraries such as TensorFlow and PyTorch, and effectively handled heterogeneous datasets—an essential feature for real-world deployment in sectors like finance and IoT.

By integrating three key FL algorithms, FedAvg, FedProx, and FedOpt, within the flower framework, the researchers successfully created models that retained high accuracy without centralized data aggregation. These algorithms allow organizations to collaborate securely, with FedAvg averaging models across clients, FedProx introducing stability in data-heterogeneous conditions, and FedOpt improving global model convergence via server-side optimization.

Does federated learning match or surpass traditional AI in cybersecurity and fraud use cases?

To assess real-world applicability, the research deployed these federated algorithms on two benchmark datasets: UNSW-NB15 (for intrusion detection) and a credit card fraud dataset. The study ran extensive simulations over 50 rounds involving 10 clients per dataset, both with and without the flower framework.

The results were decisive. With the flower framework:

• FedAvg and FedProx both achieved over 99.9% accuracy in fraud detection.
• FedOpt achieved a peak of 99.94% accuracy in both fraud and intrusion scenarios.
• Precision, recall, F1-score, and AUC scores also surpassed 90% across the board.

Without the flower framework, the same algorithms started with lower initial accuracy and slower convergence. For instance, FedOpt performed poorly in early rounds without flower, starting at just 55% accuracy, but stabilized above 99% after flower was implemented. The flower framework improved the overall model performance across all evaluated metrics, from recall and precision to convergence time.

Importantly, the flower framework offered consistency and speed. For instance, FedProx showed early stability, achieving over 99% accuracy from the initial rounds and maintaining high performance. These improvements make federated learning via the flower framework a practical, scalable option for critical applications in cybersecurity and finance.

What are the implications for future AI deployment in privacy-sensitive domains?

The study's findings highlight the transformative potential of federated learning for industries that handle sensitive, distributed data. Financial institutions, healthcare systems, and IoT ecosystems all stand to benefit from collaborative AI systems that do not require data centralization.

The flower framework demonstrated robust performance in both theory and application. Its architecture supported seamless integration with convolutional neural networks (CNNs), like the enhanced IIDNet model used in the study. The CNN model efficiently processed structured network traffic and transaction data without violating privacy standards. It was also compatible with edge devices, essential for resource-constrained IoT environments.

From a strategic perspective, federated learning reduces legal and regulatory risks. Since raw data never leaves client premises, organizations can comply with privacy laws like GDPR and India's Data Protection Act without sacrificing analytical capabilities. Moreover, federated learning ensures resilience against cyber threats such as inference attacks, data poisoning, and unauthorized surveillance.

The research also underscores the importance of personalization and inclusivity in collaborative AI systems. Federated learning allows for model customization at the client level, enabling institutions to maintain their unique data environments while contributing to a shared, global intelligence. It promotes trust, transparency, and equitable participation among stakeholders.

Looking ahead, the authors recommend further exploration of federated learning in other sectors, including autonomous vehicles, personalized medicine, and smart city infrastructure. The scalability of the flower framework also supports future innovation in real-time IoT applications, energy efficiency optimization, and decentralized edge AI systems.