The increasing scale of deep neural networks has necessitated model compression techniques, with pruning emerging as a prominent approach to reduce computational and memory costs. However, aggressive pruning introduces a critical challenge: the degradation of mathematical stability and adversarial robustness. Recent research reveals that highly pruned weight matrices tend to become ill-conditioned, exhibiting exploding condition numbers that undermine model performance and robustness . This paper proposes a condition number-aware pruning framework that explicitly preserves mathematical stability during the pruning process. We establish theoretical connections between sparsity, condition number, and local Lipschitz continuity, demonstrating that the condition number becomes the dominant factor limiting robustness in over-sparsified models . Our methodology integrates a differentiable Condition Number Constraint (CNC) with transformed sparse regularization (TSCNC) to simultaneously achieve high sparsity and well-conditioned weight matrices. Experimental evaluations on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that our approach significantly improves both standard accuracy and adversarial robustness compared to conventional pruning methods, achieving superior performance across VGG, ResNet, and WideResNet architectures.
Jeromy R, K Bhavani, K.Mohana Lakshmi et al.· International journal of com...· 0 citations
The proliferation of Internet of Things (IoT) devices has generated unprecedented volumes of data, rendering traditional cloud-centric processing paradigms inadequate due to latency constraints, bandwidth limitations, and privacy concerns. This paper presents a comprehensive mathematical framework for distributed intelligence at the edge, enabling decentralized learning across heterogeneous IoT networks. We propose a novel Federated Edge Learning (FEL) architecture that integrates software-defined networking principles with gossip-based communication protocols to facilitate collaborative model training while preserving data locality. The framework addresses critical challenges including device heterogeneity, non-independent and identically distributed (non-i.i.d.) data distributions, resource constraints, and communication efficiency. We formalize the decentralized learning problem, derive convergence bounds under heterogeneous conditions, and introduce a multi-worker selection mechanism optimized through swarm learning principles. Experimental validation using real-world IoT datasets demonstrates that our approach achieves 30-50% reduction in training latency and 35-55% decrease in energy consumption compared to conventional federated averaging methods, while maintaining competitive accuracy of 92.86% on classification tasks . The proposed framework offers a scalable, privacy-preserving solution for deploying artificial intelligence at the network edge.
N.Durga, A. Mary Posonia, Selvakumar et al.· International journal of com...· 0 citations