Content
I first encountered fastText for text classification in July of the previous year, while working on webshell detection. I ultimately used CNN text classification for that project. There was already material on fastText available at the time, so I do not cover the theory here; earlier posts address it.
Without a GPU, I decided to try fastText for text classification after wrapping up the automated threat-list work.
The full project is here, with usage instructions included. I used the public dataset from 360netlab, plus an Alexa top 100 million list.
After 250 epochs, accuracy reached 1. The model was large, however, so quantize can be used during training to compress it. This reduces accuracy: the higher the cutoff value, the smaller the model and the lower the accuracy.
Other
Last week I found a serious data leak on an internal platform and reported it directly to the project lead. A large batch of sensitive data, including training datasets from different platforms, had been exposed. The initial response was to patch only the frontend. When I checked again, path traversal still worked and it was still possible to read other users’ data without authorization. By design, non-users should not be able to download anything, and users should not be able to access one another’s datasets. After only a few sentences, he brought up the Mooncake incident and started lecturing me about values before the authorization flaw was addressed. This “technical expert”—honestly, what a joke.
Even if most people know the right values, that does not mean they will act accordingly—let alone people with ulterior motives. Elevating every issue into a question of values is not helpful.