Name
Abstract | ||
|---|---|---|
Deep learning is playing a more and more important role in our daily life and scientific research such as autonomous systems, intelligent life and data mining. However, numerous studies have showed that deep learning with superior performance on many tasks may suffer from subtle perturbations constructed by attacker purposely, called adversarial perturbations, which are imperceptible to human observers but completely effect deep neural network models. The emergence of adversarial attacks has led to questions about neural networks. Therefore, machine learning security and privacy are becoming an increasingly active research area. In this paper, we summarize the prevalent methods for the generating adversarial attacks according to three groups. We elaborated on their ideas and principles of generation. We further analyze the common limitations of these methods and implement statistical experiments of the last layer output on CleverHans to reveal that the detection of adversarial samples is not as difficult as it seems and can be achieved in some relatively simple manners. |
Year | DOI | Venue |
|---|---|---|
2018 | 10.1109/BIGCOM.2018.00021 | 2018 4th International Conference on Big Data Computing and Communications (BIGCOM) |
Keywords | Field | DocType |
deep learning, adversarial attacks, adversarial generation algorithms, easy detection | Panic,Iterative method,Computer science,Artificial intelligence,Autonomous system (Internet),Deep learning,Statistical classification,Artificial neural network,Machine learning,Scientific method,Adversarial system | Conference |
ISBN | Citations | PageRank |
978-1-5386-8022-3 | 0 | 0.34 |
References | Authors | |
0 | 3 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Feixia Min | 1 | 0 | 0.34 |
| Xiaofeng Qiu | 2 | 0 | 1.69 |
| fan wu | 3 | 7 | 7.88 |