SpamShield

SpamShield: High-Performance Multilingual Spam & Category Detection

<div align="center">

Production-grade, ultra-low-latency multilingual spam detection and 6-category classification suite powered by ONNX Runtime.

Overview • Model Variants • Benchmarks • Quickstart • Categories • NSFW Scanner

</div>

---

📋 Overview

SpamShield is an industrial-strength moderation engine built for real-time chat environments (Telegram, Discord, web applications). It couples dual-tier text analysis with computer vision media verification:

1. Binary ONNX Classifier: High-precision detection of spam vs. legitimate (ham) messages.
2. Category ONNX Classifier: Multi-class categorization across 6 distinct spam threats.
3. NSFW Vision Model: Embedded computer vision ONNX classifier for explicit media scanning.

✨ Key Advantages

• ⚡ Sub-5ms Inference: Native ONNX graph execution with zero heavy transformer overhead.
• 🔤 Unicode & Homoglyph Resilient: Built-in NFKD normalization, leetspeak decoding, character repetition suppression, and Cyrillic/Greek homoglyph transliteration.
• 🎯 Three Tuned Variants: Select between Lite, Flash (default), and Edge depending on your throughput and accuracy requirements.
• 🛡️ Zero External API Dependency: Runs completely on CPU / offline with minimal RAM footprint (<20MB).

---

🚀 Model Variants

---

📊 Performance & Benchmarks

Evaluated on global multi-lingual test sets across English, German, Russian, Spanish, Hinglish, Hindi, and Arabic.

---

🏷️ Detection Categories

The category model classifies flagged messages into 6 distinct vectors:

• 🎣 phishing — Fake login portals, credential harvesters, spoofed domain links.
• 💼 job_scams — Fake freelance/remote hiring scams, task-based pay scams, upfront deposit requests.
• 🪙 crypto — Airdrop drainers, wallet seed requests, fake trading bots, pump-and-dump signals.
• 🔞 adult — NSFW text, escort services, explicit solicitations, illicit media links.
• 📢 marketing — Unsolicited bulk promotion, channel cross-spam, aggressive affiliate links.
• 🎁 giveaway — Fake lottery rewards, gift card scams, social media reward manipulation.

---

💻 Quickstart

1. Installation

pip install onnxruntime numpy

2. Basic Inference (Python)

import numpy as np
import onnxruntime as ort

# Load binary classifier (e.g. Flash variant)
binary_session = ort.InferenceSession("Flash/binary_model.onnx", providers=["CPUExecutionProvider"])
category_session = ort.InferenceSession("Flash/category_model.onnx", providers=["CPUExecutionProvider"])

def classify_message(text: str):
    # Pass text array to ONNX string input
    input_data = np.array([[text]], dtype=object)
    
    # 1. Binary classification
    input_name = binary_session.get_inputs()[0].name
    outputs = binary_session.run(None, {input_name: input_data})
    spam_prob = float(outputs[1][0][1]) if len(outputs) > 1 else float(outputs[0][0])
    
    is_spam = spam_prob >= 0.5452  # Flash threshold
    
    # 2. Category classification if spam
    category = "ham"
    if is_spam:
        cat_input_name = category_session.get_inputs()[0].name
        cat_outputs = category_session.run(None, {cat_input_name: input_data})
        category = str(cat_outputs[0][0])
        
    return {"is_spam": is_spam, "spam_probability": spam_prob, "category": category}

# Test sample
result = classify_message("Claim 5000 USDT reward immediately by connecting your wallet here: http://fake-airdrop.xyz")
print(result)
# {'is_spam': True, 'spam_probability': 0.987, 'category': 'crypto'}

---

🖼️ NSFW Image Scanner

The repository also includes an ONNX computer vision model in nsfw/ for scanning image attachments:

• Architecture: MobileNet/EfficientNet vision backbone exported to ONNX Runtime.
• Input: Preprocessed 224x224 RGB image tensors.
• Output: Softmax probabilities across Safe vs. NSFW.

---

🔗 Dataset

• AbhinavMaurya/SpamShield-Datasets — Training corpus and benchmark splits.

---

📜 License

This model suite and codebase are released under the MIT License.