-
Notifications
You must be signed in to change notification settings - Fork 16
Home
Koorosh Aslansefat edited this page Nov 14, 2019
·
1 revision
Welcome to the SafeML wiki!
In this project, there are some assumptions explained as follows:
- In each dataset, columns with all zero, Nan or Inf values should be removed.
- Also, columns with zero variance should be removed or it can be possible to assign a small number like 1e-6 as their variance.
- Non-coherent datasets (e.g. XOR datasets) can reduce the accuracy of the algorithm.
- For datasets like Circular and Spiral one, it is suggested to convert the data to the circular coordination before the analysis.
- The proposed algorithm is not suitable to be used against adversarial attacks such as one-pixel attach and etc. An abnormal change in the field data should be in a way that affects estimated probability density, in order to detect the problem with the proposed algorithm.
Note: Using dimension reduction algorithms such as PCA and t-SNE can solve the existing limitations (assumption 1-4) of the proposed algorithm.