报告人：徐 畅  博士

报告时间：2023年9月27日（星期三）10:00-12:00

报告地点：琶洲实验室一楼会议室

报告名称：Adversarially Robust Deep Neural Networks

腾讯会议ID:915-193-719

报告人简介：徐畅博士 悉尼大学计算机科学学院副教授，专注于机器学习算法以及与计算机视觉相关的应用领域。他在知名期刊和顶级会议上发表了100多篇论文，多次获得杰出论文奖项，包括2023年AAAI杰出论文奖和2018年IJCAI杰出论文奖。此外，他还担任 NeurIPS、ICML、ICLR、KDD、CVPR和MM等会议的领域主席，以及AAAI和IJCAI的高级程序委员会成员。他曾被评为2017年IJCAI十佳杰出高级程序委员会成员，并在2022年获得IEEE T-MM杰出副编辑的荣誉称号。

报告摘要：New deep learning techniques keep improving accuracy on many benchmark tasks. However, deep neural networks’ weaknesses have been criticized for a while. As we chase higher accuracy, we must also think about balancing accuracy and robustness. In this talk,I’ll present our recent work on making neural networks more resistant to attacks. We’re asking: if we have a well-trained accurate neural network, how can we make it tougher against attacks without spending too much ? Specifically, for the latest vision transformer neural networks, how can we regain the balance between accuracy and toughness? We’ll go back to the start of network training and suggest a novel random approach. This method naturally makes the network more resistant to attacks.