Date of Award
Doctor of Philosophy
Dr. Zijiang Yang
Dr. Li Yang
Dr. Yan Cai
Concurrent bugs, deepfake detection, data race detection, sampling method, watermark embedding, invisible watermark
In the past decades, even the computer techniques have a significant improvement. The topic about security will be never out of date.
To meet with the requirement of computation, the number of CPU cores has changed from single core to multi-cores. At the same time, the multi-thread programs are also proposed to maximize the advantages of multi-core computing power. While even the performance has been improved, but it also brings some new issues which were never happened on sequential programs called current bugs. Data race is a major type of current bugs, it is happened when multiple threads access the same memory location, and at least one of them is write operation. Compare to general bugs, to detect data race is more difficult and more expensive.
Additionally, the deep learning is another hot topic area in recent years. With the improving of GPU’s performance, the neural network was deployed on GPU rather than CPU, because compare to CPU, GPU has a better computational ability on neural network. Deepfake is a new type of technique which was created based on deep learning. It is a means to swap faces realistically with a low cost in a short time. Because it is fake and a production of deep learning, so called “Deepfake”. This technique could be widely used on education, art, and entertainment area. However, it is also found in generating revenge porn, fake news, economic fraud. Because of its realistic characteristic, it is very hard to distinguish the authenticity of a picture or a video attacked by Deepfake.
A lot of detectors have been released on both races’ detection and Deepfake’s detection areas. For data races, the static analyze will have a lot false positive, while the dynamic analyze has fewer, but it will bring a huge extra overhead. To reduce the cost of dynamic analyze, the sampling strategy has been introduced, but current sampling tool reduced the overhead based on reducing the accuracy. As for the Deepfake part, some detection tools distinguish the fake materials by detecting abnormal biological information or the technical defects which have been fixed by the newer version of Deepfake strategy. The other branch is to find the consistent between frames, but this kind of method cannot be used to detect fake images. Besides, all the Deepfake detectors cannot guarantee 100% accuracy. Even with the evolution of the Deepfake, the accuracy may become lower and lower.
To address above issue, Atexrace has been presented to detect race which has the low overhead as the state-of-art sampling method and a better accuracy as the detector without sampling applied. Besides, invisible watermark embedding method to defend Deepfake attacking was proposed in this dissertation which has a 100% defense accuracy and never be out of date. And experiments result also confirmed the conclusion.
Guo, Yu, "A Comprehensive Security Approach on Data Race Detection and Deepfake Defense" (2020). Dissertations. 3649.