Biography
Biography pending.
Conference contributions 1
TOWARDS SCALABLE EDGE INTELLIGENCE:
Federated knowledge distillation for hybrid adaptive binarized neural networks
—the rapid growth of edge intelligence and Internet of Things (IoT) systems has intensified the need for deep learning models that are both lightweight and communication-efficient, especially in environments constrained by limited resources and strict privacy requirements. While Federated Learning (FL) enables decentralized training without raw data sharing, traditional FL approaches often suffer from high communication costs and degraded performance under heterogeneous, non-IID data distributions. Similarly, Binarized Neural Networks (BNNs) significantly reduce computational and memory requirements through binary weights and activations, yet they remain prone to instability and limited representational capacity in distributed optimization.. To overcome these challenges, we introduce Federated Knowledge Distillation–enhanced Hybrid Adaptive Binarized Neural Networks (FKD-HABNet), a unified framework that integrates hardware-aware binarization with communication-efficient federated optimization. In this approach, compact binarized student models collaboratively learn from globally aggregated teacher logits via Federated Knowledge Distillation, thereby reducing bandwidth requirements while enhancing generalization across diverse clients. The framework further incorporates Adaptive Long-Tailed Activation Binarization (ALAB), Pre-activation Guided Binarization (PGB), Reduced Approximate Stochastic Depth (RASD), probabilistic gradient pruning, and adversarial dual batch normalization to improve convergence stability, robustness, and feature representation. Extensive experiments on CIFAR-10 and ImageNet benchmarks using ResNet and VGG architectures demonstrate that FKD-HABNet consistently achieves superior accuracy, communication efficiency, and inference speed compared to conventional BNN and federated approaches. By significantly lowering communication overhead and memory consumption while maintaining strong performance under non-IID conditions, FKD-HABNet provides a scalable and practical solution for next-generation edge intelligence applications.