A Steganography Scheme Based On Adversarial Examples

In recent years,deep learning technology has made big development in many directions,and deep learning models have also been widely used in various research fields and deployed into people's lives.Although deep learning has achieved excellent performance on many complex problems,more and more studies have shown that deep learning models have many hidden security risks,which makes them vulnerable to attacks by malicious parties.One of the typical threats is the so-called "adversarial example".The adversarial example is obtained intentionally adding perturbation or noise by attackers,which can cause machine learning and deep learning models to produce incorrect output.If the perturbation and noise on the adversarial samples are constrained to be small enough,it will be difficult for the human and system to distinguish the adversarial example from the normal example.Adversarial examples may import a serious security threat to systems built on deep learning models,especially in applications that require high security.In this thesis,we first introduce some basic concepts in machine learning and deep learning,and summarize the research results of adversarial examples in recent years.Based on the previous literature review and summary,we will reclassify and summarize the related results based on adversarial examples.Furthermore,we propose a friendly application constructed by adversarial examples: a steganography scheme based on adversarial examples.In this scheme,we regard the adversarial nature of the adversarial examples itself as information,which can be decoded by the specific model only.We designed four feasible steganography schemes for the steganography mechanism,and demonstrated the performance of the four schemes under different constraints through experiments.At the end of the thesis,we built an experimental platform for our steganography mechanism to demonstrate a possible application of the steganography scheme in the thesis.