Multi-Class Web Attack Dataset with Iterative LLM-Generated Adversarial Payloads
Description
This dataset was created to evaluate and improve the robustness of machine learning-based WAFs and multi-class web attack classifiers. The baseline dataset (baseline_dataset.csv) contains 25,736 HTTP requests across six classes: legitimate traffic (Normal) and five injection vectors (XSS, SQLi, CMDi, LFI, and SSRF). Attack payloads were collected from public repositories and benchmarks (including OpenBugBounty, SecLists, PayloadsAllTheThings, PortSwigger, and specialized Kaggle sets) and merged with benign URLs to avoid domain bias. The dataset also includes four rounds of adversarial samples (llm_round1.csv to llm_round4.csv), totaling 4,462 attacks generated in a black-box setting using Llama 3.3 70B Instruct with prompts of increasing obfuscation. Additionally, attack sets generated by Qwen (llm_qwen.csv) and Claude (llm_claude.csv) are included for evaluating cross-model defense transferability. All CSV files share an identical structure: each row contains the HTTP fields (method, URL, payload), 56 extracted tabular features, and its class label (0: Normal, 1: XSS, 2: SQLi, 3: CMDi, 4: LFI, 5: SSRF). The complete list of features with descriptions can be found in the dataset_description.txt file.
Files
Institutions
- University of BelgradeCentral Serbia, Belgrade