Abstract:
Caching in device-to-device (D2D) communication networks is complicated due to the
dynamic and unpredictable nature of wireless environments. Limited bandwidth, high
latency, and disruptions make effective content caching and timely user request ful fillment challenging in D2D wireless networks. This manuscript proposes a federated
learning framework for edge caching in D2D wireless networks that can enhance the
efficiency of content caching and balance the trade-off between cache hit ratio and
memory use. The proposed methodology clusters devices based on their content sim ilarity using the K-means algorithm while accounting for user ratings and Euclidean
distance. Within the cluster, there is the master user equipment (MUE) and several
pieces of slave user equipment (SUE). The MUE is selected based on factors such as
willingness, signal-to-noise ratio, and battery percentage, with incentives offered by
the cellular operator. SUE devices share raw data based on content popularity and
usage patterns. The local model uses a graph convolutional gated recurrent unit that
predicts content caching through its ability to handle complex dependencies based on
the spatio-temporal features of the user devices. Federated learning facilitates global
model training without centralization of the raw data, which enhances scalability. A
proximal policy optimization algorithm determines the optimal content caching for
each device, allowing dynamic D2D environments to be effectively handled. Simula tion results demonstrate that the proposed solution yields strong caching performance
by reducing the average delay and improving the overall offloading probability.
