Live demo: frost.moosha.org
A browser-only proof-of-concept of FROST: Fingerprinting Remotely using OPFS-based SSD Timing (paper, Weissteiner et al.). From plain JavaScript, the page times random reads against a large file in the Origin Private File System (OPFS) and plots your SSD's access latency live. When anything else on the machine touches the same SSD (an app launching, a website loading), contention pushes those latencies up. The spikes are a side channel that leaks what you're doing.
Everything runs locally in your browser and measures only your own machine. No measurement data leaves the page, and there is no backend storage.
This is an unaffiliated reproduction for education and research. It is not the authors' code.
- Live SSD-contention monitor. Creates a probe file in OPFS (size is up to you, 1–8 GB), runs timed random reads, and charts the latency trace with live stats (median, p95, read rate, spikes over 1 ms) and a spike-filtered overlay.
- "Last read off the SSD" view. Hex dump and grayscale swatch of the bytes just read. They come from the page's own random-filled probe file. OPFS is sandboxed, so the signal is the read latency and never anyone's data.
- Website fingerprinting. A closed-world version of the paper's attack. You calibrate by capturing a few traces each of YouTube, Wikipedia, Amazon, and example.com. Then the page guesses which of them you open in another tab. Self-test mode opens one at random and identifies it, so you can see it isn't rigged.
- Open frost.moosha.org in a desktop browser. The Environment panel should show Cross-origin isolated: yes and OPFS available: yes.
- Pick a probe size larger than your free RAM (so reads go to the SSD instead of the page cache) and click Start. Wait for "Preparing probe file… 100%".
- Launch an app or load a heavy website and watch the trace spike.
- For fingerprinting, capture each site about 3 times, then try Self-test or Guess what I open.
- When you're done, click Stop, then Delete probe file to free the disk space.
Accuracy is rough in a casual setting and well below the paper's controlled-lab F1 scores. JavaScript can't reset the OS page cache, so a site's signal is strongest the first time it's opened. It also only works when the site's disk activity hits the same SSD as the probe file. Calibration lives in the tab and is never stored or sent anywhere.
Browsers coarsen performance.now() unless the page is cross-origin isolated. Without it,
the timer is too coarse for the SSD signal. The bundled server sets these headers on every
response:
Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp
Opening public/index.html directly over file:// will not work. If a reverse proxy or
CDN sits in front of the server, make sure it passes these headers through unchanged.
Requires Node.js 22+. There are no dependencies, so npm install isn't needed.
npm start # http://localhost:8080
PORT=3000 npm startOr with Docker:
docker compose up -d --build # binds 127.0.0.1:8013npm testThe unit tests (node --test) cover the ring buffer, statistics, spike filter, hex dump,
trace classifier, and server (headers, 404, path-traversal guard, /src route). The parts
that depend on timing, OPFS, and the canvas vary by hardware, so they're checked by hand with
the steps in Try it.
server.js zero-dependency static server; sets COOP/COEP, serves public/ and /src
public/
index.html, style.css the page
app.js UI, live chart, stats, and fingerprinting capture controller
probe-worker.js Web Worker that owns all OPFS I/O and the timed read loop
src/ pure logic shared by the browser and the Node tests
ring-buffer.js, stats.js, spike-filter.js, hexdump.js
fingerprint.js trace binning and the lag-tolerant template-matching classifier
test/ node:test unit tests
docs/superpowers/ design specs and implementation plans
frost.pdf the original paper
The demo covers the paper's core measurement primitive, live visualization, and a per-visitor, closed-world website-fingerprinting demo that uses template matching. It leaves out the paper's CNN classifier, open-world detection, application fingerprinting, and covert channel.
The attack is from FROST: Fingerprinting Remotely using OPFS-based SSD Timing by Hannes Weissteiner et al. (TU Graz). All credit for the research goes to the authors.