Author ORCID Identifier

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Zhipeng Cai

Second Advisor

Wei Li


In the era of Deep Learning, users are enjoying remarkably based on image-related services from various providers. However, many security issues also arise along with the ubiquitous usage of image-related deep learning. Nowadays, people rely on image-related deep learning in work and business, thus there are more entries for attackers to wreck the image-related deep learning system. Although many works have been published for defending various attacks, lots of studies have shown that the defense cannot be perfect. In this thesis, one-pixel attack, a kind of extremely concealed attacking method toward deep learning, is analyzed first. Two novel detection methods are proposed for detecting the one-pixel attack. Considering that image tempering mostly happens in image sharing through an unreliable way, next, this dissertation extends the detection against single attack method to a platform for higher level protection. We propose a novel smart contract based image sharing system. The system keeps full track of the shared images and any potential alteration to images will be notified to users. From extensive experiment results, it is observed that the system can effectively detect the changes on the image server even in the circumstance that the attacker erases all the traces from the image-sharing server. Finally, we focus on the attack targeting blockchain-enhanced deep learning. Although blockchain-enhanced federated learning can defend against many attack methods that purely crack the deep learning part, it is still vulnerable to combined attack. A novel attack method that combines attacks on PoS blockchain and attacks on federated learning is proposed. The proposed attack method can bypass the protection from blockchain and poison federated learning. Real experiments are performed to evaluate the proposed methods.


File Upload Confirmation