Executive Overview: The Zero-Trust Browser Imperative

The modern browser is the primary execution runtime for the enterprise, with over 75% of knowledge work occurring directly inside web tabs. However, traditional perimeter firewalls and centralized security gateways are structurally blind to client-side threat execution: generative zero-day spear-phishing, ephemeral credential harvesting portals, DOM-based form hijacking, and Unicode homograph spoofing occur milliseconds before network telemetry reaches cloud analyzers.

Guardian Angel introduces a paradigm shift: a 100% on-device, multi-tiered security engine that intercepts, disambiguates, and mitigates web-based threats in real time without transmitting a single byte of user browsing telemetry to external cloud endpoints.

+1,200%
AI Phishing Velocity

Annual surge in automated, generative spear-phishing lures and deceptive brand portals.

0 Bytes
Remote Exfiltration

100% on-device local execution; zero browsing history or payload data leaves the browser.

$4.88M
Average Breach Cost

Global enterprise financial exposure per compromised employee credential (IBM Cost of Breach Report).

The Inverted Pyramid: Multi-Tier Hybrid Security Architecture

Traditional browser extensions suffer from a crippling trade-off: lightweight regex/keyword filters trigger unmanageable false positive rates (over 50%), while deep neural networks introduce unacceptable tab rendering latency (>200 ms). Guardian Angel solves this via an Inverted Pyramid Cascade that combines probabilistic hashing, sub-linear frequency sketching, and contextual small language models (SLMs):

Guardian Angel: Inverted Pyramid Multi-Tier Cascade mermaid 
graph TD
    Page[Web Page Content & Target Link] --> T1["Tier 1: O(1) Dynamic Bloom Filter (0.08ms)"]
    T1 -- "Exact Blacklist Match" --> ThreatFast["🚨 Instant Threat Interception"]
    T1 -- "No Exact Match" --> T2["Tier 2: Discriminative Count-Min Sketch (0.50ms)"]
    T2 -- "Clean Lexical Profile (< 15.0)" --> SafeFast["✅ Fast-Pass Safe Exit (34% of cases)"]
    T2 -- "Ambiguous / High Keyword Density" --> T3["Tier 3: MobileBERT SLM Arbiter (19.6ms)"]
    T3 -- "P(Threat) >= 0.50" --> ThreatFinal["🚨 Threat Context Flagged"]
    T3 -- "P(Threat) < 0.50" --> SafeFinal["✅ Safe (CMS False Alarm Overridden)"]
Syntax & Lint Status Checking...
Tier 1: Dynamic Bloom Filter (Exact Signature Match)
Zero-Latency O(1) Blacklist Interception

Evaluates domain and URL signatures in 0.08 ms using a bit array with salted SHA-256 hash functions. Provides mathematical guarantees of zero false negatives for known blacklist feeds (PhishTank, OpenPhish, URLhaus) with bounded false positive probability (p ≤ 10-4).

Tier 2: Discriminative Count-Min Sketch (N-Gram Frequency Matrix)
Sub-Linear Streaming N-Gram Triage Filter

Computes 1-gram, 2-gram, and 3-gram frequencies across threat sketches in 0.50 ms. Incorporates discriminative ingestion where threat n-grams increment counters (+5) and benign n-grams apply negative penalties (-3), fast-passing clean traffic while escalating ambiguous text to Tier 3.

Tier 3: MobileBERT SLM Arbiter (Contextual Deep Learning)
On-Device WebAssembly SIMD Neural Reasoning Engine

Executes Int8 dynamically quantized MobileBERT (25.3 MB) in 19.6 ms inside a WebExtension Manifest V3 Offscreen Worker via WebAssembly SIMD (ort-wasm-simd.wasm). Acts as the supreme arbiter to disambiguate false alarms and catch zero-day social engineering.

Multi-Scenario Real-World Dataset Strategy & Boundary Curation

A machine learning model is only as effective as its boundary calibration. Guardian Angel is trained across four specialized real-world corpora spanning over 5,000 curated samples to master the precise failure modes where simpler filters fail:

Multi-Scenario Dataset Lifecycle & Component Assignment mermaid 
graph LR
    subgraph "Real-World Benchmark Feeds"
        D1["Phishing Emails (puyang2025, zefang-liu)"]
        D2["Spam & Card Scams (Enron, Deysi, SMS)"]
        D3["Strange URLs (pirocheto, alexkstern, Mitake)"]
        D4["Benign Prose (20Newsgroups, Civil Comments)"]
    end
    D1 & D2 & D3 & D4 --> Miner["Boundary Miner & Curator (70/30 Stratified Split)"]
    Miner --> T1_Sync["Tier 1: Bloom Filter Domains"]
    Miner --> T2_Sync["Tier 2: Discriminative CMS Tables"]
    Miner --> T3_Train["Tier 3: MobileBERT Focal Loss Training"]
Syntax & Lint Status Checking...
Scenario Dataset Benchmarks Used Threat Patterns & Edge Cases Addressed
1. Phishing Emails & Spoofed Senders puyang2025/seven-phishing-email-datasets, zefang-liu/phishing-email-dataset Credential harvesting portals, urgent account suspension notices, display name spoofing, and authentic corporate IT communications.
2. Spam & Financial Card Scams SetFit/enron_spam, Deysi/spam-detection-dataset, ucirvine/sms_spam Unsolicited Visa/Mastercard reward offerings, pre-approved credit lures, sweepstakes winnings, loan advance-fee fraud, and genuine workplace emails.
3. Strange Links & Malicious URLs pirocheto/phishing-url, alexkstern/phishing_urls, Mitake/PhishingURLsANDBenignURLs Number-stuffed subdomains, DGA-generated hostnames, brand names in URL paths/subdomains, raw IP hosts, and abnormal high-risk TLDs.
4. High-Purity Benign Baseline fetch_20newsgroups (medical, science, technical, commerce), google/civil_comments Suppresses false alarms across legitimate technical articles, software documentation, and corporate portals to guarantee sub-0.5% False Positive Rates.

Mathematical Optimization: Asymmetric Focal Boundary Loss

Under standard Cross-Entropy Loss, easily separable examples dominate backpropagation gradients, causing the model to underperform on subtle boundary cases. Guardian Angel implements Asymmetric Focal Loss with a focusing factor of γ = 2.0 and class weights α = [1.6, 1.0]:

&mathcal;LFocal(pt) = -αt · (1 - pt)2.0 · log(pt)

By exponentially damping the gradient from well-classified easy samples, over 95% of backpropagation updates are concentrated on ambiguous boundary samples—specifically teaching the SLM to rescue authentic IT/banking alerts while detecting stealthy zero-day social engineering.

Empirical Benchmarks & Multi-Tier Cascade Results

Rigorous benchmarking across a 30% holdout test set demonstrates the dramatic precision multiplier achieved by cascading probabilistic data structures with an SLM arbiter:

0.08 ms
Tier 1 Bloom Filter

Instant domain blacklist matching via bitarray hashing.

0.50 ms
Tier 2 Count-Min Sketch

Sub-linear streaming n-gram frequency matrix scoring.

15.15 ms
Full Cascade Latency

Weighted average end-to-end response time across all web traffic.

99.14%
Cascade Precision

True positive accuracy across multi-scenario threat evaluations.

0.36%
False Positive Rate

Only 1 false alarm per 281 safe pages (down from 50.89% on standalone CMS).

25.3 MB
Runtime Footprint

Int8 quantized ONNX student model fitting comfortably within browser extension memory budgets.

Metric Standalone Tier 2 (CMS) Standalone Tier 3 (SLM) Full Three-Tier Cascade (BF + CMS + SLM)
False Positive Rate (FPR) 50.89% (keyword trap) 2.49% 0.36% (1 FP / 281 Safe) 🔥
Precision 49.47% 94.44% 99.14% 🔥
True Positive Rate (Recall) 92.11% 78.29% 75.66%
Accuracy 65.59% 90.76% 91.22%
CMS False Alarms Rescued 0 — 139 safe texts correctly overridden to SAFE
Fast-Path Pass-Through — — 34.0% (147 clean texts exited in < 1 ms)

Integrated Client-Side Heuristic Defense Modules

Operating synchronously alongside the multi-tier machine learning pipeline, Guardian Angel integrates five zero-latency heuristic inspectors:

Real-Time Client-Side Heuristic Defense Flow mermaid 
graph LR
    DOM[Active Tab DOM & Network Events] --> Form[Cross-Origin Form Action Inspector]
    DOM --> Homograph[IDN Homograph & Confusable Normalizer]
    DOM --> Hidden[SpamAssassin Hidden DOM Text Detector]
    DOM --> Obfuscation[JS Obfuscation & Hex Density Scanner]
    DOM --> Favicon[Favicon MurmurHash3 Brand Checker]
Syntax & Lint Status Checking...
Real-Time Heuristic Security Inspectors
Zero-Latency Deterministic Threat Interception
  • Cross-Origin Form Action Inspector: Flags login and password forms that submit credentials to unauthorized third-party origins.
  • IDN Homograph & Confusable Normalizer: Decodes internationalized domain names (Punycode) and maps Cyrillic/Greek lookalike glyphs to ASCII canonical forms.
  • SpamAssassin Hidden DOM Text Detector: Calculates visibility ratios, invisible CSS positioning (left: -9999px), and zero-font-size text traps.
  • JS Obfuscation & Hex Density Scanner: Analyzes script tags for packing signatures (Packer, JSFuck, Obfuscator.io) and excessive hex density.
  • Favicon MurmurHash3 Brand Matching: Computes 32-bit MurmurHash3 digests of page favicons to detect brand impersonation on unverified domains.

Deploy & Install Guardian Angel

Guardian Angel is packaged as a standard Manifest V3 Chrome Extension bundle containing all pre-compiled WASM binaries, quantized ONNX models, and probabilistic filter tables. Download the deployment bundle below to load in Developer Mode:

⬇️ Download Guardian Angel Extension Bundle (v7.0 ZIP - 23.06 MB)
Quick Installation Steps:
  1. Download and unzip the guardian_angel_extension_bundle.zip archive.
  2. Open Google Chrome / Chromium and navigate to chrome://extensions.
  3. Enable Developer mode in the top-right corner.
  4. Click Load unpacked and select the unzipped guardian_angel_extension directory.
  5. Pin the Guardian Angel shield icon to your toolbar to view real-time tab risk scoring!