Phish

What it is

Phishing via SMS (smishing) and encrypted messaging channels has evolved dramatically. Attackers no longer rely exclusively on poorly translated, obvious typos; generative tooling now crafts convincing institutional alerts that impersonate banking portals, electric utilities, postal carriers, and telecom operators.

Phish is a dual-track security and applied-AI project that tackles this problem directly on the client. Rather than shipping users' sensitive personal text messages and two-factor codes to an external cloud API, Phish runs inference locally in the user's browser using an 8-bit quantized DistilBERT ONNX safetensors model (v3vishal/phish-v0.1) via WebAssembly, paired with client-side OCR for SMS screenshots and sender identifier verification.

// pure NLP misses spoofing.
sender headers don't lie.

Why Multi-Point Verification Matters

A common mistake in AI security tools is relying solely on text NLP. Consider this realistic smishing scenario:

"Your SBI Bank account has been blocked today due to pending KYC update. Please update your PAN immediately to avoid permanent suspension."

A standalone text classifier might recognize the urgency, but text alone cannot confirm illegitimacy. What exposes the attack conclusively is the multi-point context:

• Sender Identifier Integrity: Legitimate banks utilize carrier-registered alphanumeric headers (e.g. TRAI standard VM-HDFCBK or AX-SBIINB). When an official bank alert arrives from a random 10-digit personal phone number (+91 98231...), it is an immediate critical red flag.
• Domain & URL Forensics: Attackers conceal destinations with free URL shorteners, dynamic DNS, bare IP links, or high-abuse TLDs (e.g. .xyz, .top), or distribute Android banking trojans disguised as direct .apk download links.
• DistilBERT ONNX Semantic Intent: Analyzes sentence structure, manipulative coercion, and linguistic framing to score the probability that the text aims to induce panic or offer fraudulent rewards.
• Client-Side Screenshot OCR: Users don't need to copy-paste messy text. Dragging a screenshot of the conversation runs in-browser optical character recognition (Tesseract.js) to isolate the sender header and message payload automatically.

Model Architecture & Quantization

The v0.1 model is fine-tuned on specialized smishing corpora and exported to ONNX format:

• Base Architecture: DistilBertForSequenceClassification (6 Transformer layers, 768 hidden dimension, 12 attention heads, ~66M parameters).
• ONNX Graph & 8-bit Quantization: Converted from PyTorch safetensors into an int8 quantized ONNX graph (model_quantized.onnx, ~67 MB), enabling smooth WebAssembly execution in the browser via Transformers.js.
• Zero-Knowledge Privacy: Neither the screenshot, message text, nor extracted phone numbers leave the browser. Inference runs entirely inside the client memory sandbox.
• Instant Heuristic Layer: If the 67 MB ONNX model is still downloading or the user is operating in offline/low-bandwidth conditions, an instant neural heuristics fallback scores the message immediately.

Roadmap & v0.2 Directions

• Multilingual Smishing Support: Fine-tuning on Hindi, Hinglish, and regional Indian language smishing datasets to detect local scam variations.
• Visual UI Overlay Analysis: Expanding OCR into computer vision heuristics that evaluate UI styling in screenshots to detect spoofed banking application interfaces.
• WebGPU Acceleration: Leveraging WebGPU execution providers for sub-50ms token inference latency on supported hardware.

Links & Artifacts

Hugging Face Model ↗ Launch in Crypt Sec-Ops ↗ GitHub Repository ↗ ← All projects