学术讲座通知:Trustworthy Video Understanding Deep Learning Systems
主讲人：Prof. Yuan Hong
会议时间：2023/3/15 9:00-11:00
会议地址：#腾讯会议：717-841-619
主持人：喻鹏副教授
主办单位：北京邮电大学网络与交换技术国家重点实验室、计算机学院
Abstract:
Deep learning has achieved many big innovations in all industries and significantly impacted our daily lives. However, deep learning can also result in severe security and privacy risks to users and service providers. In this talk, I will present our recent works on trustworthy deep learning systems. First, we design and implement the first black-box attack framework that generates universal 3-dimensional (U3D) perturbations to subvert a wide variety of video understanding deep learning systems. The new U3D attack is easy-to-launch, universal, transferable, and human-imperceptible. It can also bypass the state-of-the-art defense methods. Such novel attack motivates the video recognition systems to build and integrate more robust deep learning models. Second, we design and implement the first cryptographic system (Crypto3D) for private and efficient deep learning based on spatial-temporal features extracted from videos. Crypto3D significantly outperforms existing systems (extensible to inferences on 3D features) on execution time: 186.89x vs. CryptoDL (3D), 63.75x vs. HEANN (3D), 61.52x vs. MP-SPDZ (3D), 45x vs. E2DM (3D), 3.74x vs. Intel SGX (3D), and 3x vs. Gazelle (3D); and on accuracy: 82.3% vs. below 56% for all of them.
Bio:
Yuan Hong is an Associate Professor in the Department of Computer Science and Engineering at the University of Connecticut. He received his Ph.D. degree from Rutgers University in 2014. His research interests primarily lie in the fields of differential privacy, secure multiparty computation, applied cryptography, trustworthy machine learning, and cyber-physical systems security and privacy. He is a recipient of the NSF CAREER award, Cisco Research Award, and Meta (Facebook) Research award finalist. His research contributions are published in prestigious security and data science venues such as IEEE S&P (Oakland), CCS, PETS, KDD, VLDB, ECCV, AAMAS, EMNLP, ICDE, CIKM, EDBT, ICDCS, TDSC, TIFS, TOPS, and TKDE. He regularly serves as the TPC or senior PC member for relevant top conferences such as CCS, USENIX Security, PETS, NeurIPS, ICML, CVPR, ICCV, ECCV, KDD, and AAAI. His research is supported by numerous NSF and AFOSR awards.
