Reproduce “LOKI: Proactively Discovering Online Scam Websites by Mining Toxic Search Queries” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pujan Paudel, Gianluca Stringhini
LOKI ranks search queries by how likely their results are to contain scam websites, helping downstream detectors discover candidates beyond those reported through reactive channels. It uses search-result information to train a query-only model, and its in-the-wild search identified 52,493 oracle-classified scam sites among 271,161 returned websites—reported as a 20.58-fold improvement over comparison strategies.
Key insights
- Search-result descriptions provide useful training-time context: the teacher generally beats a query-only baseline, while the distilled student also improves on that baseline despite not seeing SERPs at inference.
- Simple indicators such as query length, competition, intent, or particular query segments vary in usefulness by category and do not offer a consistently strong cross-category rule.
- Query ranking complements rather than replaces site classification: LOKI prioritizes searches that can supply better candidates to downstream detection systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.