🛡️ Silver Guard — Indian Multilingual SMS Scam & Threat Classifier (v2.0)

Silver Guard is an ultra-lightweight (<15 MB RAM, 56 KB graph + 15.3 MB embedding table), zero-latency edge model and 3-stage cascaded inference engine engineered specifically for detecting SMS scams, phishing lures, cyber extortion, and commercial spam across the Indian telecommunications ecosystem.

  • GitHub Repository: tanishqmudaliar/SilverGuard
  • Model Type: FastText Subword Character N-Gram Embeddings + Tabular Structural Domain MLP (Dual-Branch Early Fusion)
  • Taxonomy: 4-Class (0: Ham, 1: Promotional, 2: Commercial Spam UCC, 3: Scam / Phishing)
  • Runtime: ONNX Runtime (CPU) — Edge ready for Android smartphones (ARM Cortex-A53/A55)
  • Supported Languages: English, Hindi (हिंदी), Marathi (मराठी), Tamil (தமிழ்), Telugu (తెలుగు), Bengali (বাংলা), Gujarati (ગુજરાતી), and Romanized Hinglish.

🏷️ 4-Class Classification Taxonomy

Silver Guard decouples unsolicited commercial advertising from malicious fraud:

Class ID Semantic Class Typical Senders Key Characteristics & Behavior
0 Safe Ham Personal contacts, banks (AD-HDFCBK-T), government (JX-ISATHI-G), service alerts (MH-AIRTEL-S) OTPs, balance debits, family chats, courier delivery updates. Threat score: $\approx 0.0$.
1 Promotional Registered TRAI DLT senders with -P suffix or registered retail brands (JK-SMRTBR, JM-AJIOIN, DOMINO) Legitimate e-commerce sales, discount codes, retail offers. High commerciality, zero security threat.
2 Commercial Spam (UCC) Unregistered 10-digit mobile numbers (+919876543210) Unsolicited cold marketing (real estate, pre-approved personal loans, insurance leads). Do-Not-Disturb violation, but no credential theft or malware.
3 Scam / Phishing Unknown phone numbers, spoofed display headers Digital arrest extortion, law enforcement intimidation (CBI/Customs), fake KYC deactivation, lottery fraud, malicious APK downloads.

🧠 Neural Architecture: FastText Subword Domain Fusion

Silver Guard uses a dual-branch hybrid architecture designed for zero out-of-vocabulary (OOV) errors and ultra-low RAM consumption on low-end mobile devices:

  1. Subword Linguistic Branch: FastText character n-grams ($n \in [3, 5]$) mapped via deterministic 32-bit CRC hashing into 60,000 buckets, embedded using PyTorch nn.EmbeddingBag(60000, 64, mode="mean"). Provides complete invariance to Hinglish phonetic misspellings, leetspeak (p0lice, k-y-c), and regional slang.
  2. Domain Feature Tabular Branch: 12 structural telecom indicators (URL counts, link shorteners, direct IP links, APK downloads, callback phones, currency symbols, urgency cues, promo keywords, uppercase ratio, digit ratio, length, Indic script codes) processed through a 2-layer MLP (12 -> 32 -> 32, ReLU).
  3. Early Fusion & Head: Embeddings are concatenated into a 96-dimensional representation (64 + 32 = 96) and routed through a classification head (96 -> 4) trained with class-balanced focal weights.

Performance & Edge Hardware Comparison

Metric Silver Guard v2.0 (FastText Domain Fusion) Legacy Baseline (MobileBERT FP32)
Model Size 56 KB graph + 15.3 MB weights 105 MB
Inference Latency (CPU) ~0.8–1.2 ms 18–25 ms
Peak RAM Footprint < 15 MB (Immune to Android Go LMKD kill) ~120 MB (High OOM risk)
Hinglish / Leetspeak Handling Character n-grams (Zero OOV) High OOV ([UNK] tokens)
Indic Matras Preservation 100% (Canonical NFC Normalization) Erased by standard ASCII tokenizers

📊 Multi-Source Training Dataset (24,514 Unique Samples)

The model was trained on a multi-source unified dataset compiled and deduplicated by build_unified_dataset.py using MinHash 5-shingle Jaccard similarity ($J \ge 0.85$ deduplication threshold) and canonical NFC normalization:

Dataset Source Origin / Platform Ingested Samples Class Contribution Key Value to Model
Atherion 30K SMS Kaggle (atherionsystems/indian-scam-sms-30k) 8,599 Ham (5,977), Scam (1,851), Spam (771) Indian financial scams, lottery lures, transactional alerts
UCI SMS Spam Collection Hugging Face (uciml/sms-spam-collection-dataset) 5,009 Ham (4,825), Scam/Spam (184) Clean conversational English personal chats (eliminates personal text false alarms)
Personal Real Inbox SMS Device Export (data/rawdata.json & contacts.csv) 3,992 Ham (3,822), Promo (170) Ground-truth Indian DLT carrier traffic (TRAI headers, OTPs, delivery alerts)
ScamShield Multilingual Kaggle (siddharthrane07/scamshield-multilingual) 4,331 Phish / Scam (2,904), Ham (1,427) Multilingual smishing, credential phishing, malicious URL attacks
JuniorAlive India SMS Hugging Face (junioralive/india-spam-sms-classification) 1,983 Ham (1,483), Spam / UCC (500) Real Indian telemarketing, cold commercial offers
Curated Indic Regional Synthetic / Handcrafted Templates 600 Promo (558), Spam (42) Marathi, Hindi, Telugu, Tamil, and Gujarati retail promotions
Total Clean Dataset Unified & Deduplicated 24,514 Ham: 17,013 | Scam: 4,889 | Spam: 1,884 | Promo: 728 Full 4-Class Indian Telecom Coverage

Note on Hugging Face YAML Metadata: The datasets: YAML frontmatter tag above lists uciml/sms-spam-collection-dataset and junioralive/india-spam-sms-classification because Hugging Face YAML validation requires valid Hugging Face Hub dataset identifiers. The additional sources (Atherion 30K and ScamShield) originate from Kaggle, and the personal ground-truth inbox export is preserved privately with PII redaction.


📦 Hosted Files

  • silverguard_fasttext_fusion.onnx: Dual-branch ONNX compute graph (~56 KB).
  • silverguard_fasttext_fusion.onnx.data: Weights tensor data containing the subword embedding dictionary (~15.3 MB).
  • model_config.json: Hyperparameters, class labels, and supported languages metadata.

🚀 Quickstart: Running Inference in Python

import numpy as np
import onnxruntime as ort
import re
import zlib
import unicodedata

# 1. Load ONNX Session
session = ort.InferenceSession("silverguard_fasttext_fusion.onnx")

# 2. Preprocess Text (FastText Subwords + Hashes)
def compute_subwords(text: str, vocab_size: int = 60000) -> np.ndarray:
    text = unicodedata.normalize("NFC", text.lower())
    words = re.findall(r"\w+", text)
    hashes = []
    for w in words:
        padded = f"<{w}>"
        for n in (3, 4, 5):
            for i in range(len(padded) - n + 1):
                ngram = padded[i:i+n]
                h = zlib.crc32(ngram.encode("utf-8")) % vocab_size
                hashes.append(h)
    return np.array(hashes if hashes else [0], dtype=np.int64)

# 3. Extract 12 Domain Features
def extract_tabular_features(text: str) -> np.ndarray:
    feats = np.zeros(12, dtype=np.float32)
    feats[0] = len(re.findall(r"https?://", text))      # url_count
    feats[1] = 1.0 if "bit.ly" in text or "tinyurl" in text else 0.0 # shortener
    feats[2] = 1.0 if ".apk" in text.lower() else 0.0   # apk link
    feats[7] = 1.0 if re.search(r"\b(offer|discount|sale|off)\b", text, re.I) else 0.0 # promo
    feats[8] = sum(1 for c in text if c.isupper()) / max(len(text), 1) # upper ratio
    feats[10] = len(text) / 160.0                        # length
    return feats.reshape(1, 12)

# 4. Run Model Prediction
text = "Dear customer, your SBI account will be blocked today. Click http://bit.ly/sbi-kyc to update PAN."
subwords = compute_subwords(text)
offsets = np.array([0], dtype=np.int64)
tabular = extract_tabular_features(text)

inputs = {
    "text_indices": subwords,
    "text_offsets": offsets,
    "tabular_features": tabular
}

logits = session.run(["logits"], inputs)[0][0]
probs = np.exp(logits - np.max(logits)) / np.sum(np.exp(logits - np.max(logits)))

labels = {0: "Safe Ham", 1: "Promotional", 2: "Commercial Spam (UCC)", 3: "Scam / Phishing"}
pred_class = np.argmax(probs)

print(f"Prediction: {labels[pred_class]} (Confidence: {probs[pred_class]:.4f})")
# Output: Prediction: Scam / Phishing (Confidence: 0.9995)

📜 License & Citation

This model is licensed under the MIT License.

@software{silverguard2026,
  author = {Tanishq Mudaliar},
  title = {Silver Guard: Multilingual Edge SMS Threat & Scam Classifier for Indian Telecommunications},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/tanishqmudaliar/SilverGuard}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Evaluation results