2026· IEEE Transactions on Network and Service Management· Vol 23, pp. 8352-8367· 0 citations· 50 references
TL;DR
DeSPU is proposed, a decentralized sparse personalized federated unlearning framework under a practical deletion-window protocol that estimates a target-class direction with a lightweight dual-model mechanism, applies semantic-layer adaptive linear unlearning to suppress target knowledge, and performs topology-aware local repair regularized by cached neighbor snapshots collected during routine peer-to-peer training.
Abstract
Edge intelligence-enabled Internet of Things (Edge-IoT) increasingly relies on decentralized on-device learning to reduce cloud dependence and keep locally collected IoT data on terminal devices. However, deployed models may still memorize privacy-sensitive class-level knowledge and must support deletion requests after training. This paper studies class-level federated unlearning in a serverless Edge-IoT workflow where each IoT client maintains a sparse personalized IoT-local model and exchanges model information only with neighbors. We propose DeSPU, a decentralized sparse personalized federated unlearning framework under a practical deletion-window protocol. DeSPU estimates a target-class direction with a lightweight dual-model mechanism, applies semantic-layer adaptive linear unlearning to suppress target knowledge, and performs topology-aware local repair regularized by cached neighbor snapshots collected during routine peer-to-peer training. The online unlearning path does not store historical gradients and does not introduce additional post-request model communication beyond routine decentralized exchanges; the one-time $\beta $ calibration scan is reported separately, and the calibrated $\beta $ can be reused under the same configuration. To propagate deletion status beyond the requester, DeSPU also attaches a metadata-only forgetting flag to those routine exchanges and invalidates affected cached neighbor References at receiving clients. Experiments on CIFAR-10, MNIST, CIFAR-100, and CINIC-10 show near-zero target-class accuracy with high retained accuracy, up to $\sim 4.7\times 10^{3}$ speedup over full retraining on CIFAR-10, and membership-inference ROC-AUC close to random guessing. Additional analyses over seeds, requester clients, target classes, cache staleness, sparsity, communication, and hardware cost further demonstrate the practicality and limitations of DeSPU for privacy-compliant Edge-IoT deployments. The project link is available at https://github.com/aaaM2002/UN_CODE
Experiments show that H-FedSN reduces communication costs by up to 477 times compared to baseline methods while maintaining high accuracy, making it well-suited for hierarchical FL in IoT deployments.
Jie-Chao Gao, Yuan-Gang Li, Jie Wang et al.· npj Artificial Intelligence· 2 citations
Federated Learning (FL) enables distributed model training without sharing raw data, making it particularly suitable for Internet of Things (IoT) environments. However, FL remains vulnerable to poisoning attacks, where malicious clients manipulate local updates to corrupt global model optimization. Existing defenses ma...
Jin Meng, S. Amitrano, Marzia Canzaniello et al.· ACM Interational Symposium o...· 0 citations
MAEFL is proposed, a Multi-Agent Explainable Federated Learning framework that couples a federated demand-prediction surrogate with per-node multi-agent Q-learning policies and a permutation-importance explainability layer, enabling privacy-preserving, adaptive, and interpretable resource orchestration across heterogen...
Soumadip Mondal, Sweta Chopdar, K.Vengatesan et al.· 2026 10th International Conf...· 0 citations
Federated Learning (FL) enables collaborative model training across distributed clients while preserving data locality and reducing the need to share raw data. However, deploying FL over non-terrestrial networks (NTNs), including satellite communication systems, introduces challenges such as propagation delay, intermit...
Reemiah Muneer Alotaibi, Seid Mehammed Abdu, Md Nasre Alam et al.· Discover Computing· 0 citations
Traditional cloud-centric IoT analytics suffers from unacceptable delays, large bandwidth consumption, and potential privacy leakage; standard Federated Learning (FL) approaches addresses privacy leakage but introduce communication overhead and ignore the timing constraints imposed by heterogeneous IoT edge networks. W...
T. Krishna, S. Ul Huq, N. S. Kumar et al.· 2026 10th International Conf...· 0 citations
DPGR-GCAM, a communication-efficient federated domain generalization framework that employs a dual-parameter-efficient fine-tuning (PEFT) architecture, and introduces the Gradient-Conflict Aware Modulation (GCAM) mechanism, to mitigate client drift.
Yuan-Kun Xia, Hui Wang, Ling Yuan et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.