Research map / Large language models / AI and society
AutoMalDesc: Large-Scale Script Analysis for Cyber Threat Research
Alexandru-Mihai Apostu, Andrei Preda, Alexandra Daniela Damir, Diana Bolocan, Radu Tudor Ionescu, Ioana Croitoru, Mihaela Gaman
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 |
|---|---|---|---|
| LoRA fine-tune on 900 expert seed samples (Base to V1) | Llama-3.3-70B-Instruct · Malware detection (Batch) | Accuracy (%): 52.7 -> 77.9 | Improved |
| LoRA fine-tune on 900 expert seed samples (Base to V1) | Llama-3.3-70B-Instruct · Malware detection (average) | Accuracy (%): 83.1 -> 89.3 | Improved |
| add self-training on 101K filtered pseudo-labeled scripts (V1 to V2) | Annotator V1 · Malware detection (average) | Accuracy (%): 89.3 -> 91.5 | Improved |
| add self-training on 101K filtered pseudo-labeled scripts (V1 to V2) | Annotator V1 · Malware detection (Batch) | Accuracy (%): 77.9 -> 82.4 | Improved |
| fine-tune on seed data (Base to V1) | Llama-3.3-70B-Instruct · Language detection (average) | Accuracy (%): 93.7 -> 96.8 | Improved |
| add self-training (V1 to V2) | Annotator V1 · Language detection (average) | Accuracy (%): 96.8 -> 97.1 | Improved |
| add self-training (V1 to V2) | Annotator V1 · Language detection (PowerShell) | Accuracy (%): 92.7 -> 91.8 | Hurt |
| add self-training (V1 to V2) | Annotator V1 · Summary quality (human preference) | Win rate (%): 46.26% -> 44.85% | No clear effect |
| add self-training (V1 to V2) | Annotator V1 · Summary quality (LLM judges) | Win rate (%): Claude 3.7 V2 52.43%; Phi-3.5-Mini V1 56.22% | Mixed |
| fine-tune on seed data with minimal prompt vs prompted base (Base to V1) | Llama-3.3-70B-Instruct · Summary quality (LLM judges) | Win rate (%): Phi-3.5-Mini 55.35% V1; GPT-5 51.39% Base | Mixed |
| self-trained V2 vs prompted base | Annotator V2 · Summary quality (LLM judges) | Win rate (%): Claude 3.7 53.95% V2; Mixtral 54.67% Base | No clear effect |
| raise confidence threshold from 90% to 99% at temperature 0.8 | Pseudo-label filtering · Label retention | Retention (%): 86.49 -> 65.76 | Hurt |
Closest papers by meaning
- A Large-Scale Multi-Source Dataset Linking Hacker Community Discourse to the CVE Vulnerability Lifecycle (NeurIPS 2026 · similarity 0.84)
- Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models (ICLR 2025 · similarity 0.84)
- Is Function Similarity Over-Engineered? Building a Benchmark (NeurIPS 2024 · similarity 0.84)
- Virus Infection Attack on LLMs: Your Poisoning Can Spread "VIA" Synthetic Data (NeurIPS 2025 · similarity 0.84)
- KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards (NeurIPS 2026 · similarity 0.84)
- HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models (ICLR 2025 · similarity 0.83)
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.
