NovaSentinel-ChainFed: A Blockchain-Based Secured Federated Deep Learning Framework for Trustworthy IoT Communication

Abstract The proliferation of Internet of Things (IoT) devices has created unprecedented challenges in network security, particularly in detecting and mitigating distributed denial-of-service (DDoS) attacks. Traditional centralized intrusion detection systems face scalability limitations and privacy concerns when deployed across distributed IoT infrastructures. This article introduces \textit{NovaSentinel-ChainFed}, a framework that integrates blockchain-based coordination with federated deep learning to establish trustworthy IoT communication channels. The proposed architecture comprises three synergistic components: (i) the NovaSentinel-ID intrusion detection pipeline employing an ensemble of gradient boosting methods (Random Forest, Extra Trees, Gradient Boosting, XGBoost, and LightGBM) unified through an elastic-net regularized logistic regression meta-learner; (ii) a federated learning simulation across five IoT gateways executing ten communication rounds; and (iii) a blockchain-inspired logging layer for immutable audit trails using SHA-256 hash chaining. Central to the framework is a deep learning-based trust score mechanism that dynamically evaluates client reliability using eight behavioral features, including update norms, cosine similarity with peer updates, local validation performance, attack rate exposure, and historical trust trajectories, thereby enabling adaptive traffic policy decisions. Experimental evaluation on the CICDDoS2019 dataset demonstrates the effectiveness of the framework, achieving an accuracy of $77.08\%$, macro-precision of $85.79\%$, macro-recall of $84.42\%$, a macro F1-score of $85.09\%$, and a ROC-AUC of $93.74\%$. The system successfully classifies four traffic categories (BENIGN, DrDoS\_DNS, DrDoS\_LDAP, and DrDoS\_MSSQL) with per-class F1-scores ranging from $76\%$ to $95\%$. The trust mechanism exhibits adaptive behavior across federated rounds, with initial heuristic-based scores converging towards learned behavioral patterns after trust model training. The blockchain layer maintains verified integrity through 11 blocks recording 60 transactions across all rounds. Collectively, these results establish NovaSentinel-ChainFed as a viable solution for secure, privacy-preserving intrusion detection in distributed IoT environments.