2 ms·
DarkBERT: A Language Model for the Dark Side of the Internet
- jruohonen 3y agoInteresting work but with plenty of caveats: 1. The whole rationale of classifying a category of a site is dubious. More like an usual computational exercise. 2. The ransomware leak site detection is decent, but I'd expect HUMINT to perform much better for years to come. 3. The threat detection performance is close to random and hence useless (F1 around 53-54). The sample is also based on RaidForums alone and hence very small. 4. As they note, a restriction to English is a further limitation. While I would have accepted the paper, I am not sure whether there is anything novel here.