Random Polynomial Neural Networks: Analysis and Design

Huang, Wei; Xiao, Yueyue; Oh, Sung‐Kwun; Pedrycz, Witold; Zhu, Liehuang

doi:10.1109/tnnls.2023.3288577

Cited by 1 publication

(1 citation statement)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For example, given the local update gradients (or local models), a malicious server can infer whether a specific data point belongs to the training set of a particular client [12]. Moreover, an attacker in the open public environment can precisely reconstruct small batches of training samples [13][14][15], if the updated gradients are intercepted and the local model architectures are known. Such attacks pose a server privacy threat to FL systems.…”

Section: Introductionmentioning

confidence: 99%

Towards Privacy-Preserving Federated Neuromorphic Learning via Spiking Neuron Models

Han,

Fu,

Zhang

2023

Electronics

View full text Add to dashboard Cite

Federated learning (FL) has been broadly adopted in both academia and industry in recent years. As a bridge to connect the so-called “data islands”, FL has contributed greatly to promoting data utilization. In particular, FL enables disjoint entities to cooperatively train a shared model, while protecting each participant’s data privacy. However, current FL frameworks cannot offer privacy protection and reduce the computation overhead at the same time. Therefore, its implementation in practical scenarios, such as edge computing, is limited. In this paper, we propose a novel FL framework with spiking neuron models and differential privacy, which simultaneously provides theoretically guaranteed privacy protection and achieves low energy consumption. We model the local forward propagation process in a discrete way similar to nerve signal travel in the human brain. Since neurons only fire when the accumulated membrane potential exceeds a threshold, spiking neuron models require significantly lower energy compared to traditional neural networks. In addition, to protect sensitive information in model gradients, we add differently private noise in both the local training phase and server aggregation phase. Empirical evaluation results show that our proposal can effectively reduce the accuracy of membership inference attacks and property inference attacks, while maintaining a relatively low energy cost. blueFor example, the attack accuracy of a membership inference attack drops to 43% in some scenarios. As a result, our proposed FL framework can work well in large-scale cross-device learning scenarios.

show abstract