4 in 10: the leak study
We re-ran a 2023 study on every Umbra and ERC-5564 payment ever made on Ethereum. 38.97% of withdrawn stealth addresses were traced to their owner.
Findings
- 25,645 stealth addresses on Ethereum, every Umbra and ERC-5564 payment up to block 26,127,110.
- 38.97% of withdrawn stealth addresses are traced to their owner by H1 or H2.
- 12.12% of those first paid in 2025 are still traced: one in eight.
- 9,709 traced withdrawals come from one habit: withdrawing to the address that registered the stealth keys.
- 45.68% of single withdrawals fall into collector clusters, where one wallet pulls in many stealth addresses.
Background
A stealth address hides who received a payment. In 2023, Kovács and Seres measured how well that holds in practice for Umbra, the stealth payment system ScopeLift launched on Ethereum in 2021 (arXiv:2308.01703). On Ethereum they linked 48.51% of withdrawn Umbra payments to their recipient: 4,696 of 9,680.
Since then, ERC-5564 has standardised stealth addresses, and its Announcer went live on Ethereum in 2024. We asked whether the picture changed.
What we measured
We rebuilt their method as an open-source scanner, ran it on the full history of the chain and added one heuristic of our own. The definitions are in Method. In short:
| Heuristic | What it links | |
|---|---|---|
| H1 | Funds withdrawn to an address that registered stealth keys | stealth address → identity |
| H2 | Funds withdrawn back to the address that sent them | stealth address → identity |
| H3 | Several stealth addresses withdrawn to the same address | stealth addresses → one owner |
| H4 | Withdrawals sharing a rarely used priority fee | stealth addresses → one owner |
| H5 | Gas for a token-only stealth address paid by its owner's side (ours) | stealth address → identity |
Results
| Scope | Stealth addresses | Withdrawn | Linked, H1 or H2 | Share |
|---|---|---|---|---|
| Paper, Ethereum 2023 (reported) | 9,680 | 4,696 | 48.51% | |
| Umbra as of 2023-07-01, our rerun | 12,667 | 12,419 | 6,872 | 55.33% |
| First paid since 2023-07-01 | 12,978 | 12,578 | 2,760 | 21.94% |
| All, Umbra and ERC-5564 | 25,645 | 25,017 | 9,749 | 38.97% |
By year of first payment
| Year | Withdrawn | Linked | Share |
|---|---|---|---|
| 2021 | 782 | 556 | 71.1% |
| 2022 | 7,415 | 4,598 | 62.01% |
| 2023 | 8,039 | 2,687 | 33.42% |
| 2024 | 4,078 | 1,324 | 32.47% |
| 2025 | 3,820 | 463 | 12.12% |
| 2026 to date | 883 | 121 | 13.7% |
Each heuristic, all data
| Heuristic | Result |
|---|---|
| H1, single withdrawal | 9,709 |
| H2, single withdrawal | 426 |
| H3, collector pattern | 11,425 addresses in 3,786 clusters, the largest of 53; 45.68% of single withdrawals |
| H4, unique priority fee | 1,024 transactions in 463 groups, from 22,616 eligible |
| H5, gas funding | 0 of the 2 token-only addresses that paid their own gas |
What it means
- The math is not the weak point. Every link above comes from what happened after the payment.
- It is getting better. The traced share fell from 71% in 2021 to 12% in 2025.
- One in eight is still too many. A tool called private should not leave the most common leak to the user. That is why the ephemeral client blocks or warns on these habits instead of documenting them.
Our rerun versus the paper
On the paper's own window our pipeline finds 55.33%, against the paper's 48.51%. It also counts more withdrawn addresses, 12,419 against 9,680. Our definitions are written out in full in Method, so the difference can be checked line by line; we do not claim the paper is wrong.
Limits
- These are lower bounds. Real analysts also use timing, amounts and off-chain data.
- Implementations that never announce on-chain are invisible to the scan.
- Exchange deposit addresses count as ordinary recipients.
Data and privacy
The scanner writes aggregate counts only. It never outputs which address was linked to whom, and we do not publish such lists. Run it yourself: Reproduce the results.