The Ledger
Researchers Test Proof-of-Reserves for Finding Exchange Wallets
A paper by six Dunamu researchers uses proof-of-reserves disclosures to label Ethereum exchange addresses and tests a machine-learning classifier on 3 million addresses. The model scores strongly on held-out exchanges but shows uneven recall on separate wallet sets, limiting its use to screening rather than definitive attribution.

A classifier built by six Dunamu researchers scored a 0.9594 macro PR-AUC when tested on Ethereum exchange addresses it had not seen during training, but its recall fell to 59.90% on a separate group of smaller exchange wallets. The results, in a paper submitted to arXiv on October 30, show both the promise and uneven reach of using on-chain behavior to identify centralized exchange (CEX) addresses.
The researchers use proof-of-reserves (PoR) disclosures to label known exchange wallets, then train a model to screen other addresses based on transaction patterns. Their strongest result comes from a structured leave-one-exchange-out test; the weaker figure comes from a different set of addresses and a different measure. The figures are not directly comparable, but together they show why the authors describe the method as a screening tool rather than a definitive way to assign ownership.
PoR labels anchor a 3 million-address dataset
The team assembled 3,033,857 Ethereum addresses: 1,580,155 labeled as exchange wallets and 1,453,702 across five non-exchange categories. Those categories cover bridge and lending protocol users, decentralized-exchange liquidity pools, Ethereum Name Service registrants and foundation addresses. Exchange labels came from public PoR disclosures collected from 15 platforms.
The exchange side of the dataset is heavily concentrated: Binance accounts for 1,468,505 of the 1,580,155 exchange addresses. The researchers grouped 11 exchanges with relatively few disclosed addresses into an “Others” category rather than using them in the main held-out-exchange test. They drew Ethereum records through block 25,046,605, dated May 7, 2026.
From those records, the team calculated 58 features covering activity such as gas use, ERC-20 transfers, internal transactions, counterparties and ETH flows. A validation-only selection process narrowed the model’s input to 17 features. The paper tests LightGBM against logistic regression across observation windows from 30 to 730 days.
Longer histories lift the held-out score
In the main evaluation, the researchers withheld one exchange at a time—Binance, Bitget, OKX or Korbit—so addresses from that exchange did not appear in model development. With 17 features and a 730-day lookback, LightGBM reached a macro PR-AUC of 0.9594, compared with 0.7697 for logistic regression. Macro PR-AUC summarizes precision and recall across the held-out exchange tests; it is not the share of all wallets correctly identified.
Observation history mattered. With only four selected features and a 30-day lookback, LightGBM’s macro PR-AUC was 0.6003; extending that window to 730 days raised it to 0.9052. At the longest window and full feature set, recall ranged from 90.14% to 96.28% across the four held-out exchanges when the decision threshold was calibrated to a 1% false-positive rate on validation data.
“These findings support interpretable CEX address screening for blockchain monitoring, while the released feature dataset enables reproducible follow-up research.”
Seongchan Park et al., Dunamu
External wallets expose uneven coverage
The researchers also tested two additional sets that were not used for model development or testing. One contained 209 PoR-linked addresses from smaller exchanges; the other contained 656 exchange addresses manually collected from top-holder lists for 45 tokens on Etherscan and Arkham. The latter labels were less direct than the PoR disclosures, the paper says, but included wallet-role annotations.
At the validation-calibrated 1% false-positive threshold, the model recovered 90.34% of addresses in the 656-address external set, but only 59.90% of the smaller “Others” group. Results also varied by exchange: recall was 98.68% for Gate and 84.31% for Coinbase, against 56.25% for KuCoin and 50.00% for Bitkub. The model performed better on hot and deposit wallets than on cold or unspecified addresses in the same evaluation.
The authors stop short of definitive attribution
Gas behavior, ERC-20 activity and internal transactions contributed most to the model’s predictions. Removing gas-use and fee features caused the largest drop in macro PR-AUC, while removing ERC-20 activity or internal-transaction features also reduced performance. The authors caution that these signals describe model behavior within their dataset, not inherent traits shared by every exchange.
The study covers Ethereum mainnet and depends on publicly disclosed exchange addresses and five selected non-exchange groups. The authors say that adding other types of wallets, including custodians, payment processors and institutional accounts, could better represent non-exchange activity. They also note that performance varies by exchange and wallet role, and say the classifier should help prioritize addresses for further investigation—not replace curated exchange labels.
Sources
- Original source · Preprint On-Chain Behavioral Signatures for Classifying Centralized Exchange Addresses on Ethereum Using Proof-of-Reserves arXiv:2610.41888
Related stories
- Nethermind Researchers Test EIP-8141’s Mempool DoS Limits
- Researchers Find 4.67M Malicious EIP-7702 Delegations
- Lean Proof Caps EIP-7002’s Block Call at 210,000 Gas
- Scroll replaces mainnet proof verifier after vulnerability report
- Blockaid says attacker rotated Gonka bridge key, drained reserves and minted 22 million WGNK