講座名稱：Secure Distributed Computation for Federated Learning and Blockchains

講座人：Songze LI

講座時間：5月29日15:00

地點：北校區(qū)新科1012

講座人介紹：

李松澤是香港科技大學(xué)（廣州）物聯(lián)網(wǎng)方向和人工智能方向的助理教授，也是香港科技大學(xué)計算機(jī)科學(xué)系的助理教授。在2020年加入香港科技大學(xué)之前，他曾在斯坦福大學(xué)擔(dān)任過兩年的區(qū)塊鏈研究員。于2011年在紐約大學(xué)獲得學(xué)士學(xué)位，并于2018年在南加利福尼亞大學(xué)獲得博士學(xué)位，均為電子工程專業(yè)。目前主要研究領(lǐng)域包括人工智能安全與隱私、安全的多方計算及隱私以及區(qū)塊鏈的安全性和可擴(kuò)展性。目前擔(dān)任兩個國際信息理論和信息安全期刊的客座編輯，還擔(dān)任了14個人工智能和通信領(lǐng)域的頂級國際期刊和會議的TPC成員。在NeurIPS-20 Workshop on Scalability，Privacy，and Security in Federated Learning上獲得了最佳論文獎。2017年高通創(chuàng)新獎學(xué)金的入圍者。

講座內(nèi)容：

Security and privacy issues are becoming increasingly critical to the development anddeployment of modern information systems. This talk provides an overview on ourrecent results of developing secure, robust, and efficient distributed computationprotocols, for the application of federated learning (FL) and blockchains. Two thingswill be discussed.Federated Learning:We expose the vulnerability of FL models via developing a novel backdoor attack that significantly improve the backdoor durability over SOTA; we develop novel secure model aggregation protocols that respectivelyminimizes the computation and communication overhead, while providinginformation-theoretic privacy for clients’ data; we also develop secure model/embedding aggregation protocols for horizontal and vertical FL scenarios, simultaneously achieving lossless performance and perfect privacy in presence of client dropouts.Blockchains:We propose coded Merkle tree, a novel cryptographic accumulator that allows a blockchain light client to securely and efficiently verify the availability of a block; we also develop PolyShard, a novel blockchain sharding protocol that leverages coded computation to simultaneously scale the storage efficiency, throughput, and security of block verification, in presence of adaptive adversaries.

主辦單位：空天地一體化綜合業(yè)務(wù)網(wǎng)全國重點實驗室