When: Wednesday, October 24, 2018
Where: PGH 563
Time: 11:00 AM

Determining Affine Equivalence of Boolean Functions

Speaker: Luis Moraes

Abstract and Bio TBA

Detecting Intruders and Preventing Hackers from Evasion by Tor Circuit Selection

Speaker: Zechun Cao

The widely-used Tor network has become the most popular anonymous network that supports circuit-based low latency internet connections. However, recent security breach incidents reveal SSH have been used to launch attacks by malicious users. Although a server-side blocking mechanism which can identify SSH connections individually has been proposed, we have found that it is restricted to certain Tor circuit protocol versions and not for all SSH protocol implementations. The prior method is based on the difference of latency in the Tor network which may be subject to hacker manipulation by circuit selection in the Tor network. In this paper, we first present a set of attributes that can be used to detect SSH connection through Tor for all SSH handshake between client and server, by observing the network packets exchanges of the SSH protocol. In the second half of this paper, we show that the geographical location of the nodes in Tor circuit has an impact on the effectiveness of our metrics. If hackers know our detection algorithm, they may be able to evade the detection. We demonstrate the effectiveness of our attacks detection by analyzing multiple Tor circuit selections. Finally, we identify and evaluate our detection algorithm and demonstrate that our algorithm achieves 98% accuracy under the most stringent condition.

Bio:

Zechun Cao received his Master's degree in Computer Science from the University of Houston in 2013. After working as Software Developer at Baylor College of Medicine for two years, he is currently pursuing a Ph.D. degree in Computer Science at the University of Houston. His research interests encompass network security, intrusion detection, and machine learning.

Real-time Facial Expression Reconstruction and Transformation from Video

Speaker: Luming Ma

Realistic facial expression creation and transformation has been a long-standing problem in computer graphics and computer vision. Thus far, popular approaches usually require a driving source or the combination of multiple ones, such as capturing a subject’s performance and then transferring it to virtual faces, and speech-driven facial animation. However, these methods only provide a way to drive the face to follow the performed expressions, and do not provide the flexibility to synthesize new facial expressions on top of the original, such as being happier or being angry instead of neutral while speaking. In addition, the transferring approaches usually break the synchronization between the face reenactment and audio from the source video and thus are unsuitable for speech video. In this talk, we present a complete pipeline to real-time transform the source expression of the subject in an input monocular RGB video clip to a user-specified target expression and then photo-realistically re-render the same performance but with the target expression. The generated facial expression sequence is temporally dynamic, coherent, and lip-synchronized to the source audio.

Bio:

Luming Ma is a fourth year Ph.D student in Department of Computer Science at University of Houston. His research focus on computer graphics especially on facial reconstruction, expression transformation and transferring on monocular RGB camera. He has rich industrial experience in online game development.

Satirical News Detection and Analysis

Speaker: Fan Yang

Despite news satire is for entertainment purpose, people can get deceived and spread news satire as true news if they fail to recognize the satire. Therefore, satirical news detection is important to prevent misleading information over the Internet. Existing text classification methods only consider document-level features for satire detection, but we observe the satirical cues have only existed in certain paragraphs and sentences. To reveal the satire, we analyze news articles into paragraphs and sentences, and propose hierarchical neural network to incorporate features at different levels, so that each paragraph and each sentence can be weighted differently based on the satirical degree. Experiments demonstrate that our method achieves substantial improvement over baselines. Based on the satirical degree obtained via our model, we further investigate a thorough analysis to understand the satire.

Bio:

Fan Yang is a fourth-year Ph.D. student in the Department of Computer Science at the University of Houston, advised by Dr. Arjun Mukherjee. Fan is interested in deep learning and natural language understanding, with a particular focus on detecting misleading information.