When I research the ransomware detection with adversarial examples, I extracted the dataset from the strings based on cuckoo dynamic analysis. Those datasets are based on the count of each string. So I generated the adversarial examples based on those datasets, while not binary.
When it comes to this question, I am not sure whether it is necessary or I was in the wrong direction.
For adversarial images examples, it is easy to just change one pixel. But for malware/ransomware binary, how to achieve that part?
Dear Dr Tan ---here u are concerning adversarial binaries in a Cyber security Set up:
https://arxiv.org/pdf/2007.02407.pdf
Generative adversarial network (GAN) is one of the most promising methods of unsupervised learning in recent years. GAN works via the adversarial training concept and has shown excellent performance in the fields of image synthesis, image super-resolution, video generation, image translation, etc.
Article Quantum Generative Adversarial Network: A Survey
For the question posed in the final line of your post have a look at :
https://arxiv.org/pdf/1710.08864.pdf
Conference Paper Adversarial Examples for Malware Detection
I trust the input helps:
Regards--
Dear Dr Durrani，
The first and last references are really helpful. They gave me some clues for my questions. For the second one, to be honest, it is what I implemented before. I originally thought there was a difference between adversarial images and binaries. It sounds like I was wrong. Thanks a lot
- Badges
- Science topic