Research map / Large language models / Language, safety and interpretability
VeriDark: A Large-Scale Benchmark for Authorship Verification on the Dark Web
Andrei Manolache, Florin Brad, Antonio Barbalau, Radu Tudor Ionescu, Marius Popescu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on Dark-Web-related VeriDark data rather than PAN2020 large | BERT authorship verification · VeriDark AV datasets | avg.: 66.9 -> 77.2; 66.1 -> 84.6; 68.2 -> 88.1 | Improved |
| Train on the test-matched dataset rather than a different VeriDark dataset | BERT authorship verification · DarkReddit+, SilkRoad1, and Agora AV tests | avg.: 77.2; 84.6; 87.3 | Improved |
| Train on All VeriDark datasets rather than only a different single dataset | BERT authorship verification · Cross-dataset AV tests | avg.: 74.7; 83.5; 88.1 | Improved |
| Add an Other class using PAN examples during training | DarkReddit+ author identification · DarkReddit+ test examples, including Other | accuracy: 84.6 -> 86.5 | Improved |
| Add an Other class using PAN examples during training | DarkReddit+ author identification · ClearReddit test examples, including Other | accuracy: 81.2 -> 80.4 | Hurt |
| Add an Other class from a similar VeriDark source | DarkReddit+ author identification · DarkReddit+ test examples | accuracy: 84.6 -> 83.7; 84.6 -> 83.5; 84.6 -> 80.3 | Hurt |
| Add an Other class from a VeriDark source | DarkReddit+ author identification · ClearReddit test examples | accuracy: 81.2 -> 75.0; 81.2 -> 77.7; 81.2 -> 73.9 | Hurt |
| Add reverse chunk pairs during evaluation (ablation) | BERT authorship verification · DarkReddit+ AV test | avg.: 76.6 -> 77.2 | Improved |
| Add reverse chunk pairs during evaluation (ablation) | BERT authorship verification · SilkRoad1 AV test | avg.: 85.0 -> 85.5 | Improved |
| Add reverse chunk pairs during evaluation (ablation) | BERT authorship verification · Agora AV test | avg.: 87.6 -> 88.0 | Improved |
| Pretrain on PAN before VeriDark fine-tuning (ablation) | BERT authorship verification · DarkReddit+ AV test | avg.: 77.2 -> 77.3 | No clear effect |
| Pretrain on PAN before VeriDark fine-tuning (ablation) | BERT authorship verification · SilkRoad1 AV test | avg.: 84.6 -> 84.2 | No clear effect |
Closest papers by meaning
- A Synthetic Dataset for Personal Attribute Inference (NeurIPS 2024 · similarity 0.83)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.82)
- CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark (AAAI 2025 · similarity 0.81)
- T-NET: Weakly Supervised Graph Learning for Combatting Human Trafficking (AAAI 2024 · similarity 0.81)
- STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings (ICML 2025 · similarity 0.81)
- SafeArena: Evaluating the Safety of Autonomous Web Agents (ICML 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
