Skip to content

AI_ARCHITECTURE.md

Latest commit

 

History

History
125 lines (95 loc) · 7.28 KB

AI_ARCHITECTURE.md

File metadata and controls

125 lines (95 loc) · 7.28 KB
Raw

AI Architecture & Multi-Modal Threat Pipeline

1. Multi-Modal Pipeline Overview

GhostNet AI operates a multi-modal artificial intelligence pipeline designed to inspect, score, explain, and reconstruct cyber threats across five distinct attack surfaces:

  1. Natural Language Messages: SMS smishing, WhatsApp lures, Telegram extortion, and spear-phishing emails.
  2. Hyperlinks & Lookalike Domains: Typosquatting URLs, brand lookalikes, IP-literal URLs, and obfuscated redirect gateways.
  3. Screenshots & Visual Evidence: Fake banking portals, counterfeit payment receipts, and QR code phishing traps.
  4. Synthetic Audio & Voice Calls: Deepfake audio impersonation, AI voice cloning, and robotic IVR fraud calls.
  5. Decentralized Threat Indicators: Community threat telemetry and synchronized global blocklists.

2. Multi-Model Inference Cascade & Failover

GhostNet guarantees high availability and zero downtime through a tiered cascade:

flowchart TD
    Ingress["Threat Ingress<br/>Text / URL / Image / Audio"] --> RouteType{Vector Type?}

    RouteType -->|"Text / Message / URL"| GroqPrimary["Groq LPU: llama-3.3-70b-versatile<br/>Sub-800ms Inference Window"]
    GroqPrimary -->|"Timeout 9s / Rate Limit"| GeminiText["Google Gemini 2.5 Flash<br/>Text and Multimodal Fallback"]
    GeminiText -->|"Network Degradation / Offline"| OfflineNLP["Local Heuristic Rule Engine<br/>Regex and Social Engineering Tokens"]

    RouteType -->|"Screenshot / QR Code"| GeminiVision["Google Gemini 2.5 Flash<br/>Visual OCR and Brand MIMIC Detection"]
    GeminiVision -->|"Fallback / Offline"| OfflineOCR["Client Canvas OCR and Pattern Engine"]

    RouteType -->|"Voice / Audio Payload"| VoicePipeline["Acoustic and Linguistic Analyzer"]
    VoicePipeline --> AcousticFFT["Acoustic Wiener Entropy: fft.js<br/>Spectral Flatness and HF Energy Ratio"]
    VoicePipeline --> WhisperSTT["Groq Whisper-large-v3<br/>Speech-to-Text Transcription"]
    AcousticFFT --> VoiceScore["Combined Deepfake Threat Score"]
    WhisperSTT --> VoiceScore

    GroqPrimary --> DefenseFilter["Code-Defense Parser: safeParseVerdict"]
    GeminiText --> DefenseFilter
    OfflineNLP --> DefenseFilter
    GeminiVision --> DefenseFilter
    OfflineOCR --> DefenseFilter
    VoiceScore --> DefenseFilter

    DefenseFilter --> StandardizedCodes["10 Fixed Reason Codes Mapping"]
    StandardizedCodes --> FinalOutput["Unified Threat Verdict<br/>Score 0-100, Kill-Chain, Intent, Reason Badges"]
Loading

3. Explainable Risk Scoring & Reason Codes Defense

Traditional machine learning classifiers suffer from opaque confidence scores and prompt hallucination vulnerabilities. GhostNet enforces strict explainability through 10 Standardized Reason Codes:

Reason Code Category Plain-English Detection Criterion
URGENCY_SCARE_TACTICS Psychological Manipulation Threatening account closure, power disconnection, or immediate legal action within short deadlines (e.g., "within 24 hours").
REQUEST_OTP_PASSWORD Credential Harvesting Explicit solicitation of One-Time Passwords (OTPs), PINs, passwords, or CVVs.
PAYMENT_REDIRECT Financial Coercion Directing users to unauthorized payment gateways, gift card purchases, or reverse UPI collect requests.
SUSPICIOUS_DOMAIN Infrastructure Deception Lookalike domains, typosquats, newly registered domains, Punycode, or raw IP addresses (http://192.168.x.x).
IMPERSONATION_BRAND Identity Spoofing Mimicking recognized financial institutions, government agencies, delivery couriers, or telecom providers.
MALICIOUS_ATTACHMENT Malware Delivery Unsolicited executable files, disguised APKs (.apk), malicious macros, or invoice scripts.
UNSOLICITED_CONTACT Cold Ingress Messages received from unknown numbers, international dial codes (+92, +234), or non-consensual channels.
POOR_GRAMMAR_FORMAT Formatting Anomaly Glaring grammatical errors, awkward machine translations, irregular capitalization, or zero-width character obfuscation.
REWARD_BAIT Incentive Lure Fabricated lottery prizes, unrequested cashback, part-time YouTube review jobs, or crypto investment windfalls.
THREAT_BLACKMAIL Sextortion / Intimidation Accusations of illicit activities, fabricated webcam recordings, or ransomware extortion.

Code-Defense Parsing (safeParseVerdict)

All LLM responses pass through safeParseVerdict():

  1. Strips markdown code fences (```json ... ```).
  2. Parses raw JSON strictly.
  3. Filters the reasonCodes array against an immutable set of the 10 authorized reason codes.
  4. Drops any hallucinated or arbitrary tokens, ensuring deterministic downstream rendering.

4. Acoustic Wiener Spectral Flatness Engine

For synthetic call and voice scam detection, GhostNet utilizes a hybrid acoustic-linguistic engine:

A. Acoustic Spectral Analysis (src/lib/spectralFeatures.js)

  • Uses fft.js to perform an in-memory 512-point Fast Fourier Transform on 16kHz PCM audio waveforms.
  • Spectral Flatness (Wiener Entropy): Ratio of the geometric mean to the arithmetic mean of the power spectrum: $$\text{Flatness} = \frac{\exp\left(\frac{1}{N} \sum_{k=0}^{N-1} \ln S(k)\right)}{\frac{1}{N} \sum_{k=0}^{N-1} S(k)}$$
  • Human biological vocal tract resonances create distinct harmonic formant peaks (lower spectral flatness ~0.1 - 0.35).
  • Neural vocoders and synthetic text-to-speech generators exhibit higher high-frequency noise and flattened distributions (>0.45), indicating synthetic or voice-cloned origins.
  • Additional extracted features: Zero-Crossing Rate (ZCR), High-Frequency Energy Ratio, and Pitch Jitter Variance.

B. Linguistic Social Engineering Analysis

  • Transcribes audio in real-time via Groq Whisper-large-v3.
  • Scans the transcribed text for high-pressure extortion keywords (e.g., "digital arrest", "customs detention", "verify OTP").
  • Combines acoustic synthetic probability (40% weight) with linguistic threat severity (60% weight).

5. Threat Reconstruction™ (5-Stage Cyber Kill-Chain)

GhostNet contextualizes every attack into an interactive 5-stage progression:

[1. Ingress Vector]
      │ (Unsolicited SMS, WhatsApp message, cold IVR call)
      ▼
[2. Social Engineering Pretext]
      │ (Urgent bank KYC expiration, electricity disconnection panic)
      ▼
[3. Phishing / Trap Gateway]
      │ (Typosquatting link, QR code trap, APK download)
      ▼
[4. Credential Harvesting]
      │ (Fake NetBanking portal, OTP interception form)
      ▼
[5. Loss & Impact]
      │ (Unauthorized UPI fund transfer, account takeover, identity theft)

6. Attacker Intent Inference Taxonomy

The Attacker Intent Engine translates technical markers into plain-English adversary motivations:

  • Financial Extraction: Unauthorized fund transfers via reverse UPI collect requests or credit card fraud.
  • Credential Harvesting: Stealing authentication secrets (passwords, PINs, OTPs) for downstream account takeovers.
  • Identity Impersonation: Collecting government ID numbers (Aadhaar, PAN, SSN) to facilitate synthetic identity theft.
  • Device Compromise: Coercing victims to download hostile APKs or remote access tools (AnyDesk, TeamViewer).
  • Extortion / Blackmail: Pressuring victims into cryptocurrency transfers under duress or fabricated legal accusations.