Managing resources and executing collaborative routing in massive cognitive radio networks (CRNs) has been accomplished with the help of deep reinforcement learning (DRL). But it takes a lot of energy and more time for transmissions because it uses trial and error to interact with the surroundings. The underlying difficulty of dynamic cognitive radio networks (CRNs) is the volatility and unpredictability of spectrum availability. These shifts have occurred because of opportunistic access, which makes conventional routing and allocating resources very difficult. Our innovative Federated Deep Reinforcement Learning (F-DRL) system for Dynamic Routing with Cross-Layer Optimizer aims to tackle this issue. All of the secondary user (SU) nodes in our architecture take advantage of DRL. By taking into account the current state of the network, it smartly decides on the best next-hop and the associated distribution channel. This decision-making procedure is carried out in a multi-layer architecture. Queue length and path delay are some of the Network Layer characteristics that it incorporates. Along with that, it takes into account Physical and MAC Layer metrics like SINR, PU operation, and channel quality. In order to circumvent the drawbacks of centralized DRL, we employ a Federated Learning framework. Some of these restrictions include the fact that exchanging data in its entirety raises privacy concerns and has a significant transmission overhead. Because of this, SUs can train their local DRL models in private; only updating the central server on model updates is necessary for the global aggregate. In comparison to cutting-edge heuristic routing protocols, the suggested F-DRL-driven adaptive routing method performs far better in simulations. Additionally, it improves upon conventional routing methods that rely on DRL. It demonstrates high performance in relation to packet delivery ratio (PDR), overall delay, and network throughput. In addition, the plan offers a way to learn on a large scale. In addition, it provides a method that is crucial for widespread adaptive CRN deployments, one that preserves anonymity.

This paper presents a new robust mutual authentication scheme known as SA-SIP for the protection of Session Initiation Protocol (SIP) in VoIP systems. This scheme differs in incorporating an additional version of the Secure PAKE that makes it possible to achieve mutual authentication while avoiding storing secret information on clients. This innovation helps to solve a critical issue of physical attacks on credentials stored during the traditional authentication process. The importance of this work, therefore, is to perform enhancement to both security and performance in SIP-based VoIP systems. First, since PAKE is incorporated in the proposed scheme, some of the attacks such as man-in-the-middle, impersonation, registration hijacking, and denial of service are eliminated but at the same time, the existing security of VoIP is also improved by eliminating the vulnerability of existing systems where credentials are usually stored. Second, the scheme is proven theoretically using the formal tool, ProVerif. Finally, the result of the performance analysis is that SA-SIP still has low computational complexity which makes the proposed scheme suitable for deployment in practical systems. Thus, by providing both security and performance characteristics SA-SIP has the advantage of being an effective and high-impact change in VoIP security.

The increasing complexity of traffic patterns in modern urban environments necessitates intelligent and scalable solutions for real-time traffic forecasting within the Vehicular Ad Hoc Networks (VANETs). The proposed hybrid lightweight Deep Neural Network (DNN)–Convolutional Neural Network (CNN) framework integrates with Geographic Routing Protocols (GRP) to enable efficient traffic prediction in resource-constrained V2X environments. The model combines the spatial feature extraction ability of CNNs with the temporal learning strength of DNNs to capture the dynamic vehicular patterns. Evaluated using the TiHAN-V2X dataset, the model achieves an MSE of 0.0141, an MAE of 0.0951, and an R $$^{2}$$ of 0.9680 that surpassing state-of-the-art baselines. The lightweight design ensures low-latency inference suitable for edge devices such as RSUs and OBUs. The contributions include a scalable hybrid architecture for spatio-temporal traffic learning, integration of geographic routing for congestion-aware communication, and validation using real-world V2X data. The results demonstrate that the proposed framework enhances predictive accuracy, reduces latency, and supports intelligent traffic management in next-generation vehicular networks.

Abstract In the context of the IoT platform, the smart home represents a quintessential application scenario. Here, device-to-device (D2D) collaboration serves as the core element of its ecosystem, playing a crucial role in implementing diversified automated execution scenarios that are customized to fulfill user requirements. The progressive integration of edge computing and AI technologies has enhanced the collaboration among heterogeneous devices. Nevertheless, the conventional centralized D2D collaboration authorization decision-making supported by a single IoT Hub violates the Principle of Least Privilege (PoLP), which is a foundational design tenet that has been empirically validated as an optimal engineering practice for enhancing system security and reliability in IoT ecosystems. If there is a trade-off of PoLP violations, it fails to meet the users’ Quality of Experience (QoE). To address this issue, we propose D2DA, a distributed authorization decision-making model for smart home D2D collaboration, which constructs a distributed decision-making consensus network suitable for the edge side of smart homes by leveraging distributed ledger technology. D2DA presents a machine learning algorithm with a time complexity of O ( n ). Through this algorithm, consensus nodes can be efficiently and dynamically selected. Furthermore, D2DA ensures the security of the D2D collaboration process via wallets and hash verification. Extensive experiments conducted on a real-world smart home scenario validate that the decision-making latency of D2DA is on par with that of a single IoT Hub mode. The average latency for verifying the correctness of the newly added execution results is only 0.08% of the system time of D2DA, which is negligible.