In the evolving consumer electronics (CE) landscape, devices such as smart cameras, wearables, and consumer drones are becoming increasingly prevalent. However, these devices face significant challenges in performing image classification tasks, including limited processing power, energy constraints, and privacy concerns. To address these limitations, we propose a decentralized federated learning (DFL) framework that enables efficient, on-device image classification by using local model training and peer-to-peer parameter aggregation. This framework is deployed on the Fog-enabled UAV-as-a-Service (FU-Serve) platform, which supports real-time, distributed processing across heterogeneous CE devices. To meet the resource constraints of CE devices, we developed DFLNet, a lightweight deep learning model based on depthwise separable convolutions. To ensure scalable and low-latency model aggregation, we incorporated a sparse mesh communication topology, minimizing synchronization overhead while preserving learning quality. Furthermore, we integrated holographic counterpart technology to provide real-time virtual representations of physical devices, enhancing coordination and operational transparency. Extensive evaluations on both the AIDER and NDD disaster datasets confirm the model’s effectiveness for privacy-preserving, real-time image classification. DFLNet reduces the model size to just 0.67 MB–92.4% smaller than conventional architectures–while maintaining competitive performance, achieving 97.82% accuracy on the AIDER dataset.