報告時間：2023年08月16日（星期三）14:30-15:30

報告地點：翡翠湖校區翡翠科教樓A座1602室

報 告 人：石怡 博士

工作單位：美國弗吉利亞理工大學

舉辦單位：計算機與信息學院

報告簡介：

A wireless sensor network is deployed to monitor signal transmissions of interest across a large area. Each sensor receives signals under specific channel conditions based on its location and trains an individual deep neural network model for signal classification. To enhance accuracy, the network utilizes decentralized federated learning over a multi-hop wireless network, allowing collective training of a deep neural network for signal identification. In this approach, sensors broadcast their trained models to neighboring sensors, gather models from neighbors, and aggregate them to initialize their own models for the next round of training. This iterative process builds a common deep neural network across the network while preserving the privacy of signals collected at different locations. Evaluations are conducted to assess signal classification accuracy, convergence time, communication overhead reduction, and energy consumption in various network topologies and packet loss scenarios. The impact of random sensor participation in model updates is also considered. Additionally, we investigate an effective attack strategy that employs jammers to disrupt model exchanges between nodes. Two attack scenarios are examined: First, the adversary can attack any link within a given budget, rendering the two end nodes unable to exchange their models. Second, jammers with limited jamming ranges are deployed, and each jammer can only disrupt nodes within its range. When a directional link is attacked, the receiver node cannot receive the model from the transmitter node. We develop algorithms to select links to be attacked in both scenarios and design algorithms to deploy jammers optimally, maximizing their impact on the decentralized federated learning process. We evaluate these algorithms using wireless signal classification as the use case over a large network area, exploring how these attack mechanisms exploit various aspects of learning, connectivity, and sensing.

報告人簡介：

石怡，博士，現為美國弗吉利亞理工大學副教授，曾任美國智能自動化公司首席研究員。石怡副教授是國際知名的人工智能安全和優化領域專家，在國際著名期刊和會議上發表論文180多篇，其中單篇文章他引數超過100次的有20多篇，單篇他引次最高的超過800。石怡副教授曾兩獲無線網絡著名會議INFOCOM的最佳論文獎，分別是2008年和2011年；石怡副教授還獲得過ACM WUWNet 2014年最佳學生論文獎和IEEE HST 2018年最佳論文獎。石怡副教授擔任過多個IEEE和ACM Symposium、Track、Workshop的技術委員會主席，以及IEEE Communications Surveys and Tutorials和IEEE Transactions on Cognitive Communications and Networking的編輯。