Research Article
Detection of Trojaning Attack on Neural Networks via Cost of Sample Classification
| Input: The node sensitivity on each layer , number of layers of the model , number of neurons on each layer | | Output: Identify the abnormal nodes | | (1) | for to do: | | (2) | Sort the elements of the distribution in ascending order | | (3) | Calculate the lower quartile position in : | | (4) | Calculate the median quartile position in : | | (5) | Calculate the upper quartile position in : | | (6) | Calculate based on their positions | | (7) | Calculate interquartile range | | (8) | Calculate weak upper limit | | (9) | Calculate weak lower limit | | (10) | Calculate strong upper limit | | (11) | Calculate strong upper limit | | (12) | if | | (13) | output the node is an abnormal node | | (14) | end if | | (15) | end for |
|
