dvara 0.1.1

High-speed malicious URL detection using a Bloom Filter

dvara

High-speed malicious URL detection using a Bloom Filter. Checks 3 million URLs in 5MB of RAM.

pip install dvara

dvara check https://suspicious-site.com
🚨 MALICIOUS | urlhaus | malware_download | 2.5ms | online

dvara check https://google.com
✅ CLEAN | 0.1ms | online

---

What is dvara?

dvara is a Python CLI and library for detecting malicious URLs using a Bloom Filter — the same probabilistic data structure used internally by Chrome Safe Browsing.

It ingests threat feeds from URLhaus, PhishTank, and OpenPhish (~86,000 URLs updated daily), stores them in a 5MB Bloom Filter, and checks any URL in under 1ms — without touching a database for clean URLs.

---

Architecture

Daily ingestion job
→ Pulls URLhaus + PhishTank + OpenPhish (~86k–3M URLs)
→ Builds Bloom Filter (5.2MB, 0.1% FPR)
→ Saves to ~/.dvara/filter.bin

dvara check [url]  (online mode)
→ FastAPI backend
→ Hash URL → check 10 bit positions in Bloom Filter
→ All bits OFF → CLEAN instantly (0.1ms, DB never touched)
→ All bits ON  → query PostgreSQL confirmed_urls table
→ Found        → MALICIOUS + source + category
→ Not found    → SUSPICIOUS (false positive)

dvara check [url] --offline
→ Loads filter from ~/.dvara/filter.bin
→ Checks locally, zero network calls
→ dvara update to refresh

Two-stage design (the key insight)

Stage	What	Latency	When
1 — Bloom Filter	Redis bitstring, 10 hash lookups	0.1ms	Every request
2 — PostgreSQL	confirmed_urls table lookup	1–3ms	Only on bloom hits

Clean URLs never touch the database. False negatives are mathematically impossible.

---

Benchmarks

Metric	Result
Clean URL check	0.1ms
Malicious URL check (full pipeline)	2.5ms
URLs stored	85,976 (scales to 3M)
Filter size	5.14 MB
False negative rate	0% (guaranteed)
Target false positive rate	0.1%
Actual false positive rate	~0% at current fill

---

Installation

pip install dvara

For running the backend server

pip install dvara[server]

---

CLI Usage

Check a URL (online mode — hits API)

dvara check https://suspicious-site.com

Check a URL (offline mode — local filter, zero network)

dvara check https://suspicious-site.com --offline

Show filter and API stats

dvara stats

Update local filter cache

dvara update

Run ingestion manually

dvara ingest
dvara ingest --dry-run

---

Running the Backend

With Docker Compose (recommended)

git clone https://github.com/yourusername/dvara
cd dvara
docker compose up --build

This starts:

• FastAPI — API server on port 8000
• Redis — Bloom filter bitstring cache
• PostgreSQL — confirmed URLs table

Manually

pip install dvara[server]

# Build the filter
python -m dvara.ingestion

# Start the API
python -m uvicorn dvara.app:app --reload

---

API Endpoints

Endpoint	Method	Description
/api/check?url=...	GET	Two-stage URL check
/api/confirm?url=...	GET	Direct DB lookup
/api/stats	GET	Filter + connection stats
/api/reload	POST	Reload filter from disk
/health	GET	Health check

Example response

{
  "url": "http://110.36.95.252:49267/bin.sh",
  "result": "MALICIOUS",
  "source": "urlhaus",
  "category": "malware_download",
  "latency_ms": 2.5,
  "stage": "db",
  "checked_at": "2026-05-02T13:01:04.767776+00:00"
}

---

The Math

• n = 3,000,000 URLs, p = 0.001 (0.1% FPR)
• Bit array size: m = -(n × ln(p)) / (ln(2))² = ~43M bits = 5.2MB
• Hash count: k = (m/n) × ln(2) = 10 hash functions
• Hash algorithm: MurmurHash3 with seeds 0–9

Why Bloom Filter and not a hash set?

3M URLs in a Python hash set = 500MB+. A Bloom Filter at 0.1% FPR = 5.2MB. False positives just trigger the DB confirm — acceptable. False negatives are mathematically impossible. The Bloom Filter is the right tool.

Why Redis and not disk?

Multiple FastAPI workers need to read the same filter simultaneously. Disk requires locking. Redis bitstring is shared memory across all workers — horizontal scaling for free.

---

Threat Feed Sources

Feed	Format	URLs
URLhaus	CSV	~26,000
PhishTank	JSON (gzipped)	~59,000
OpenPhish	Plaintext	~300

---

Project Structure

dvara/
├── bloom.py        ← BloomFilter class (core)
├── ingestion.py    ← Fetch feeds, build filter
├── app.py          ← FastAPI backend
├── cli.py          ← Click CLI commands
└── config.py       ← Constants and env vars

---

License

MIT