Research map / Label-efficient and robust learning / OOD and anomaly detection
Toward More Generalized Malicious URL Detection Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies token-level data bias in malicious URL/file detection and proposes adversarial debiasing embedding (ADE) that removes class-distribution bias from token embeddings, improving out-of-domain and temporal generalization of URLNet, MalConv and LSTM detectors.
Problem
Neural malicious URL detectors trained on random splits reach high AUC but generalize poorly to out-of-domain VirusTotal daily queries and to later time periods, because low-entropy tokens are spuriously correlated with classes.
Approach
Compute per-token class-distribution matrix from training data as pseudo bias labels and train an adversarial bias predictor with gradient reversal so the token embedding loses information about those biases, minimizing mutual information between embedding and bias while keeping classification discriminative.
Key results
- On URL data, URLNet+ADE reaches 90.1 AUC vs 69.3 for plain URLNet and 75.1 for BlindEye; LSTM+ADE reaches 90.6 AUC vs 61.6 for plain LSTM.
- MalConv+ADE reaches 90.7 AUC on URL data vs 64.1 for plain MalConv, and 86.7 AUC vs 80.7 on HTML.
- Temporal sliding-window experiments show ADE-equipped URLNet consistently outperforming the original model over time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adversarial debiasing embedding | URLNet · VirusTotal daily queries, URL | AUC: 69.3 -> 90.1 | Helped |
| add adversarial debiasing embedding | Malconv · VirusTotal daily queries, URL | AUC: 64.1 -> 90.7 | Helped |
| add adversarial debiasing embedding | Malconv · VirusTotal daily queries, HTML | AUC: 80.7 -> 86.7 | Helped |
| add adversarial debiasing embedding | Malconv · VirusTotal daily queries, EXE | AUC: 67.0 -> 71.3 | Helped |
| add adversarial debiasing embedding | LSTM · VirusTotal daily queries, URL | AUC: 61.6 -> 90.6 | Helped |
| add adversarial debiasing embedding | LSTM · VirusTotal daily queries, HTML | AUC: 80.8 -> 87.9 | Helped |
| add adversarial debiasing embedding | LSTM · VirusTotal daily queries, EXE | AUC: 62.5 -> 68.6 | Helped |
Surprising findings
- Random k-fold validation with AUC above 98% on ISCX-URL-2016 coexists with ~30% AUC drops on later or out-of-domain data.
- Benign tokens such as 'facebook' and 'apple' were highlighted as malicious signals because they appear in phishing URLs present only in the malicious class.
Limitations
- URLNet is applied only to URL data because of inefficiency on longer HTML and EXE sequences.
- HTML and EXE test sets are small, so gains there are less significant.
- BlindEye is designed for image classification and is adapted to this setting.
Experimental setup
Datasets: ISCX-URL-2016 VirusTotal daily queries (URL) VirusTotal HTML crawled from URLs VirusTotal EXE
Benchmarks: ISCX-URL-2016 to VirusTotal daily queries (out-of-domain) sliding-window temporal backtest on VirusTotal URLs
Metrics: AUC balanced accuracy F1-score
Baselines: URLNet MalConv LSTM BlindEye
Builds on: BlindEye gradient reversal Integrated Gradients SmoothGrad
Closest papers by meaning
- Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy (AAAI 2025 · similarity 0.84)
- VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild (NeurIPS 2026 · similarity 0.83)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.83)
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model Accuracy (ICLR 2025 · similarity 0.82)
- Scanning Trojaned Models Using Out-of-Distribution Samples (NeurIPS 2024 · similarity 0.82)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 2025 · similarity 0.82)
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.