ADAPTIVE TRUST-WEIGHTED FEDERATED GENERATIVE DEEP LEARNING FOR PRIVACY- PRESERVING INDUSTRIAL IOT SECURITY

Industrial Internet of Things (IIoT) environments are increasingly exposed to sophisticated
cyber threats such as ransomware, data manipulation, insider attacks, and model poisoning.
Traditional centralized deep learning-based intrusion detection systems compromise data
privacy and are unsuitable for cross-industrial collaboration. Although federated learning
enables decentralized training, it typically assumes honest participants and remains vulnerable
to adversarial model updates. This paper proposes an Adaptive Trust-Weighted Federated
Generative Deep Learning framework for secure and privacy-preserving Industrial IoT
environments. The proposed system integrates generative deep learning models to simulate
zero-day industrial attack patterns and mitigate dataset imbalance without exposing sensitive
operational data. A hybrid deep neural network architecture is employed for time-series
anomaly detection in IIoT sensor streams. To enhance robustness, a novel adaptive trust
weighted aggregation mechanism dynamically assigns credibility scores to participating nodes
based on gradient deviation, historical reliability, and anomaly consistency. This approach
mitigates model poisoning attacks within federated settings while preserving data
confidentiality. The framework operates under a zero-trust security paradigm, ensuring
continuous verification of edge devices and preventing malicious model contributions.
Experimental evaluation demonstrates improved detection accuracy, reduced false positive
rates, enhanced resilience against adversarial updates, and minimized privacy leakage
compared to conventional federated and centralized approaches. The proposed model provides
a scalable, robust, and regulation-compliant security solution for next-generation Industry 5.0
ecosystems.