Process unstructured data sources, such as issues

[pombredanne]

These contain valuable data nuggets among an ocean of junk and we need to be able to find the good things there.

Some sources are:

• mailing lists such as:
  • Collect Apache Announces mailing list #100
  • Collect bugtraq mailing list #104
  • Collect data from mailing lists #108
• changelogs Parse CHANGELOGs to discover new Vulnerabilities #233
• reflogs of commit (see also the commits from vulncodedb and SAP/Eclipse steady KB)
• bug and issue trackers (such as Django, etc)
• actual description of a CVE or the text body of advisories. See Extract interesting data from CVE and other vulnerabilities body #551

We can either automate it all, but that's going to be super difficult, or rather start to craft a curation queue and parse as much as we can to make it easy to curate by humans

• Create community curation model for VulnerableCode #218

... and progressively also improve some mini AI and classification to help further automate the work.

[AyanSinhaMahapatra]

A reference: https://hal.science/hal-03430826/document

[ThePhilosopher4097]

Interested in the Project Idea...

[ThePhilosopher4097]

Interested in the Project Idea...
I think, processing of changelogs, reflogs of commits and mailing list data can be a automated

[TG1999]

Please also check: https://github.com/cve-search/git-vuln-finder

[TG1999]

https://github.com/pyupio/changelogs

[ykodwani01]

I guess the process of change logs of Apache mailing list can be automated using OpenAI' API or other open source LLMs, where we scrape the data using Selenium, feed into LLM, get the output as json format and then update the database accordingly. What is your view on that @pombredanne . #218 Can also be implemented.

[Suraj209211]

Automating the extraction of valuable information from Apache mailing list changelogs using OpenAI’s API and other tools is a great initiative and I think for the unstructured data we can focus primarily onto the Dataset for feature Engineering and classified into diverse group

Model Training: Fine-tune the selected model on a prepared dataset of CVEs in code. This will help the model learn to identify vulnerabilities in the unstructured data..... As well as we can use LoRA for the model to train

Vulnerability Detection: Use the trained model to parse through the unstructured data and identify potential vulnerabilities. This could involve using NLP techniques to understand the vulnerability descriptions and infer the vulnerable package name and versions.

Most Important Parameter to be checked is this
Text Classification: This involves categorizing text into predefined groups. [In vulnerability detection, this could be used to classify descriptions as either indicating a vulnerability or not

Information Extraction: This is the process of automatically extracting structured information from unstructured text data.

@pombredanne @AyanSinhaMahapatra

[tech-wizar]

@pombredanne I’m really interested in this problem of finding meaningful vulnerability data from noisy sources. Instead of full automation, I’d focus on building a smart curation system where parsing + human validation work together. Starting with structured extraction and a simple review queue, and later improving it using lightweight ML, can make the process scalable and much more efficient.