metadata
language:
- en
license: apache-2.0
tags:
- security
- cybersecurity
- waf
- web-application-firewall
- multi-label-classification
- injection-detection
- capec
- owasp
- distilbert
pipeline_tag: text-classification
widget:
- text: '[METHOD] GET [PATH] /?id=1%27%20OR%201=1-- [UA] Mozilla/5.0'
- text: >-
[METHOD] POST [PATH] /submit [UA] Mozilla/5.0 [CT]
application/x-www-form-urlencoded [BODY]
id=%3Cscript%3Ealert(1)%3C/script%3E
- text: '[METHOD] GET [PATH] /../../../../etc/passwd [UA] Mozilla/5.0'
- text: >-
[METHOD] GET [PATH] /blog/index.php/2020/04/04/hello-world [UA]
Mozilla/5.0
4thwall-WAF-v2.1b: Hardened Multi-Label Neural Web Application Firewall
4thwall-WAF-v2.1b is an enterprise-grade multi-label transformer model designed for real-time inspection, classification, and blocking of malicious web traffic. Fine-tuned on distilbert-base-uncased across 363,000+ real-world and honeypot HTTP requests using strict template-disjoint grouping, it eliminates structural URI leakage and achieves state-of-the-art semantic injection detection without relying on brittle regex patterns.
- Developer: 4thwall Security Team
- Base Architecture:
distilbert-base-uncased(6 layers, 768 hidden dimension, 12 attention heads) - Output Classes: 13 multi-label CAPEC categories (Normal + 12 attack vectors)
- Primary Metrics (Test Unique-Template, 29,996 novel URIs):
- Macro-F1 (Excl. 272): 0.9923
- Macro-F1 (All 13): 0.9830
- Micro-F1: 0.9912
- Exploit Catch Rate: 99.41% (22,448 / 22,582 attack rows)
- Clean Benign False Positive Rate: 0.782% (58 / 7,414)
Quickstart: Running Inference
pip install torch transformers safetensors
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
MODEL_ID = "your-username/4thwall-waf-v2.1b"
# 1. Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID, trust_remote_code=True)
model.eval()
# 2. Serialize an incoming HTTP request
raw_text = model.serialize_request(
method="GET",
path="/?id=1' OR '1'='1",
ua="Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36"
)
# 3. Predict probabilities
inputs = tokenizer(raw_text, return_tensors="pt")
with torch.no_grad():
outputs = model(**inputs)
probs = torch.sigmoid(outputs.logits)[0]
# 4. Apply calibrated decision thresholds
thresholds = model.config.thresholds
id2label = model.config.id2label
print(f"Input: {raw_text}\n")
for idx, p in enumerate(probs):
label = id2label[idx]
thresh = thresholds.get(label, 0.5)
if p.item() >= thresh:
print(f" [ALERT] {label:<38}: Score = {p.item():.4f} >= {thresh:.2f}")
Two-Tier Operational Enforcement Policy
In production, do not treat all classes as automated block targets. The model is designed to operate under a two-tier policy:
+-------------------------------------------------------------------------------+
| INCOMING REQUEST -> [METHOD] [PATH] [UA] [CT] [BODY] -> DISTILBERT INFERENCE |
+-------------------------------------------------------------------------------+
| |
v (Score >= Threshold) v (Score >= Threshold)
+------------------------------------+ +------------------------------------+
| BLOCK TIER (10 Classes) | | ALERT TIER (2 Classes) |
| SQLi (66), Code Injection (242), | | Protocol Manipulation (272) |
| OS Command Injection (88), LFI(126)| | Vulnerability Scanning (310) |
| Verb Tampering (274), SSRF (194), | | |
| Response Splitting (34), etc. | | |
+------------------------------------+ +------------------------------------+
| |
v v
[HTTP 403 FORBIDDEN] [SIEM LOG / RATE LIMIT]
(0.737% Benign Block Rate, (Zero drops on crawler bursts;
97.30%–98.82% Catch Rate) mitigates scanner reconnaissance)
Calibrated Decision Thresholds
| Index | Class Name | Calibrated Threshold ($T_j$) | Enforcement Tier | Operational Rationale |
|---|---|---|---|---|
| 0 | 000 - Normal |
0.55 | Baseline | Neutral benign traffic gate |
| 1 | 272 - Protocol Manipulation |
0.99 | ALERT | Tuned to eliminate 36,109 crawler false alarms on headerless traffic |
| 2 | 242 - Code Injection |
0.65 | BLOCK | Captures XSS and PHP execution exploits |
| 3 | 88 - OS Command Injection |
0.90 | BLOCK | Shell command execution (includes merged CAPEC-248) |
| 4 | 126 - Path Traversal |
0.90 | BLOCK | Directory traversal and /etc/passwd local file inclusion |
| 5 | 66 - SQL Injection |
0.70 | BLOCK | High recall (99.2%) with strict precision (0.996) |
| 6 | 16 - Dictionary-based Password Attack |
0.60 | BLOCK | Brute-force credential stuffing signatures |
| 7 | 310 - Scanning for Vulnerable Software |
0.55 | ALERT | Scanner reconnaissance (sqlmap, nikto, nuclei) |
| 8 | 153 - Input Data Manipulation |
0.65 | BLOCK | Parameter tampering and data sanitization bypass |
| 9 | 274 - HTTP Verb Tampering |
0.90 | BLOCK | Unauthorized HTTP verbs (TRACK, PROPFIND, PUT) |
| 10 | 194 - Fake the Source of Data |
0.90 | BLOCK | Header spoofing and SSRF proxy injection |
| 11 | 34 - HTTP Response Splitting |
0.60 | BLOCK | CRLF header injection and response splitting |
| 12 | 33 - HTTP Request Smuggling |
0.95 | BLOCK | TE/CL desynchronization attacks |
Request Serialization Specification
The model consumes raw HTTP transactions serialized using structured markup tokens:
[METHOD] {method} [PATH] {path} [UA] {user_agent} [CT] {content_type} [BODY] {body}
- Max Body Truncation: Body text is truncated to 500 characters to prevent token starvation of URI components.
- User-Agent Fallback: Unspecified/empty User-Agents default to standard browser strings (
Mozilla/5.0...) to prevent false alarms on scanning heuristics.
Dual-WAF Defense with OWASP CRS
In benchmark evaluations across 29,996 novel templates and 200 canonical textbook payloads:
- Neural Model Strength: Deep semantic understanding of evasion (whitespace-to-comment, SQL comments, inline encoding) in query parameters, paths, and POST bodies without rigid regex brittleness.
- OWASP CRS Strength: Rigorous protocol enforcement on HTTP headers and cookies (which the model serializer omits).
- Union Benchmark: Combining Model v2.1b with OWASP CRS v4 achieved a 99.57% Exploit Catch Rate on 22,582 attack rows, and 93.8% catch rate across textbook test vectors.
Citation & License
- License: Apache 2.0
- Model Checkpoint: Frozen Epoch 2 Checkpoint
v2_1b_best.pt