Farbling

Imagine wearing glasses with slightly different frames in each shop you visit. Staff in one shop recognise you all afternoon, because your frames there do not change. But if two shops compare their descriptions of you, they do not match. Tomorrow you pick new frames everywhere.

Brave farbles by default, deriving its noise from a per-session secret and the site you are on, and Floati's engine uses the same kind of randomisation (checked October 2026).

What the practice world reads

Pages in the practice world can read two lists of numbers from page.traits: canvas_pixels (24 values) and audio_samples (12 values), each between 0 and 255. A fingerprint script hashes all of Sam's traits, these two lists included, into one ID.

The three rules of good noise

1. Different per site. The noise must depend on the top-level site, so ShopA and ShopB get different readings and different fingerprint IDs.
2. Stable within a site and session. The same site must get the same noise on every page load, or its own features (such as "is this the device you signed in from?") start failing.
3. Small. Each value may move by at most 1. A shop that draws a chart checks its pixels are within 2 of what it expects; more noise than that and the chart looks broken.

Noise you can repeat: a seeded random generator

Python's random module can make a generator from a seed. The same seed always gives the same sequence of "random" numbers, which is exactly rule 2. Build the seed from the session key and the site, and you get rule 1 for free:

import random
rng = random.Random("session-41|shopa.example")   # the seed is any text you like
print([rng.choice([-1, 0, 1]) for _ in range(5)])  # e.g. [0, -1, 1, 1, 0]
rng = random.Random("session-41|shopa.example")   # same seed again...
print([rng.choice([-1, 0, 1]) for _ in range(5)])  # ...same five numbers

rng.choice(list) picks one item. To keep a value inside 0 to 255 after adding noise, clamp it: min(255, max(0, value)). max(0, v) lifts anything below 0 up to 0, and min(255, ...) pulls anything above 255 down.

Do not put the page's full address in the seed. Then every page of the same shop would get different noise, breaking rule 2. And never use Python's own hash() for this: it changes every time Python starts.

How you will know it worked

run_farbling(farble) loads each shop twice in each of two sessions, with your function changing the readings, and reports:

• "linked_pairs": pairs of shops the fingerprint tracker could join (want 0)
• "broken": shops whose chart came out wrong (want none)
• "unstable": shops that read different values on two loads in one session (want none)
• "linked_sessions": shops the tracker recognised again in the next session (want none)

An honest limit

Noise only defeats trackers that need an exact match. A determined script could try to average readings or round them away, which is why real engines also cap how precisely some readings can be taken, and make other traits uniform (Leg 2's "Defeat fingerprinting" lab). Farbling is one layer, not the whole wall.

function seededRandom(seed) {          // mulberry32: tiny, repeatable, fine for noise
  return () => {
    seed |= 0; seed = (seed + 0x6D2B79F5) | 0;
    let t = Math.imul(seed ^ (seed >>> 15), 1 | seed);
    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;
    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
  };
}
function farble(values, seed) {
  const rnd = seededRandom(seed);
  return values.map(v => Math.min(255, Math.max(0, v + Math.floor(rnd() * 3) - 1)));
}

use rand::{rngs::StdRng, Rng, SeedableRng};
fn farble(values: &[u8], seed: u64) -> Vec<u8> {
    let mut rng = StdRng::seed_from_u64(seed);
    values.iter().map(|&v| (v as i16 + rng.gen_range(-1..=1)).clamp(0, 255) as u8).collect()
}