Monero hides transaction details on the blockchain, while Tor is designed to conceal a user’s network address. But these mechanisms protect different layers, and using them together does not guarantee complete anonymity. In their study Deanonymizing Monero Transactions in Tor Network, the authors describe ProxyMark, a system that they say can link a transaction to the node that sent it to the network and, under additional conditions, to its IP address.
This is not a matter of breaking Monero’s cryptography or decrypting Tor traffic. The attack relies on how transactions are propagated and on the ability to correlate network events by timing. The full paper was published as an arXiv preprint in 2026; the authors also note that an earlier, shorter version was presented at the WWW’24 workshop.
Two layers of protection
Dandelion++ propagates a transaction in two stages. First, it is sent along a limited route—the stem phase—in which nodes forward it down a chain. Then the fluff phase begins, and the transaction spreads more widely across the network. This design is intended to make it harder to determine which node created the transaction.
Tor addresses a different problem: it routes a connection through a chain of relays to hide the user’s network address from the nodes they interact with. However, when Monero runs over Tor, it still uses its own rules for selecting peers and forwarding transactions. The researchers identify interactions between these mechanisms as a potential source of information leakage.
It is important to distinguish between two types of nodes that are easy to confuse. In the paper, “proxy nodes” are Monero outbound nodes that operate as Tor hidden services and participate in forwarding transactions. They are not Tor relays. A relay is a separate component of the Tor network through which encrypted traffic passes.
How ProxyMark tries to link a transaction to its sender
According to the authors, an outgoing transaction from a Monero node connected through Tor is initially sent to two selected outbound Monero hidden-service nodes. It can then reach ordinary nodes and spread across the open network. Transactions that a node merely forwards are handled differently from those it creates itself. The researchers argue that this difference could help identify the transaction’s source if an attacker controls a suitable peer.
ProxyMark combines this idea with two other stages. First, the system tries to determine the node’s role: whether it is a Tor client or a hidden service. To do this, the authors analyze periodic Monero Timed Sync messages and recurring data within them. The attacker then tries to influence the selection of outbound peers and proxy nodes, making it more likely that attacker-controlled nodes will be on the transaction’s path. Finally, the system tries to link a Monero node identifier to an IP address by observing network behavior at an entry Tor relay.
For this last stage, the researchers use a timing watermark: a sequence of changes in the number of messages at specified intervals. The Tor relay does not decrypt Monero traffic; instead, it looks for a matching timing pattern in the data stream. If that pattern can be matched to messages observed on the Monero node’s side, it may link the network activity to a particular connection. The paper describes different versions of this analysis for Tor clients and hidden services.
What the experiments showed
The authors tested components of ProxyMark on a Monero test network, the Monero mainnet, and the real Tor network. It is important not to interpret these results as a single test in which the researchers deanonymized an arbitrary user from start to finish: different stages were tested under different conditions, and some experiments used controlled configurations.
On the test network, the researchers ran 300 connections: 200 from hidden-service nodes and 100 from a Tor client. In this experiment, the system correctly identified the role of every connection and extracted the hidden services’ onion addresses; the authors report 100% accuracy and recall. This result applies to classification in the experiment described, not as a guarantee of the same accuracy in every real-world configuration.
Another series of experiments tested whether the attacker could occupy a target node’s outbound connections. The experiment was conducted at a reduced scale compared with the threat model described by the authors. After the target was restarted, attacker-controlled addresses occupied between 7 and 11 of its 12 outbound connections to hidden services. In a scenario where connections were periodically replaced, the attackers managed to occupy between 8 and 10 of the 10 connections in about 20 minutes. This suggests that peer selection can be influenced under the experimental conditions, but it does not, by itself, prove that the sender’s IP address was identified.
The authors also tested whether they could influence proxy selection on the Monero mainnet. With one controlled outbound node, the share of epochs in which the attacker’s node was selected as a proxy rose from 15.3% to 35.7%. Results were higher with more controlled peers: for example, with nine such nodes, the attacker’s proxy was selected in all 45 observed cases. The authors link this effect to the influence of blockchain-height data on proxy selection.
Finally, in the timing-watermark experiments, the target node used a Tor relay controlled by the researchers as its only guard. Under this condition, watermark detection achieved 100% accuracy and 93.8% recall for hidden-service nodes, and 91.4% recall for Tor clients. These figures apply to watermark detection in a controlled configuration. They do not mean that any Tor relay can see any user’s IP address and transactions.
How broad is the risk?
A full attack requires a combination of conditions. The attacker needs controlled Monero nodes that can interact with the target and influence its peer selection. Linking the activity to an IP address additionally requires a controlled Tor relay to be positioned at a suitable entry point in the connection. The authors separately discuss the likelihood of this happening and provide estimates for different infrastructure sizes, but these are calculations within their threat model—not evidence of widespread, real-time deanonymization of users.
The paper’s conclusion is therefore not that Monero has been “broken” or that Tor is useless. The authors show that network analysis can weaken anonymity at the transaction-source level, even while the transaction contents and cryptographic mechanisms remain protected. Linking a transaction to a node is not the same as revealing the parties’ addresses on the blockchain, obtaining a private key, or identifying the user.
What could be changed
The authors’ proposals are mainly aimed at developers. They consider changing Timed Sync responses so that a distinctive, recurring onion address cannot be used to classify a node so easily. They also consider changing how transactions are forwarded over Tor. The latter is particularly important: if a node cannot tell from the fact of receiving a transaction whether its peer created it, distinguishing an outgoing transaction from a forwarded one becomes more difficult.
The paper also discusses limits on the number of addresses accepted, checks on reported blockchain height during proxy selection, and limits on the frequency of service requests. These measures are intended to make it harder to influence peer selection and create timing watermarks.
ProxyMark’s main takeaway is that anonymity depends not only on cryptography and not only on Tor. Network-interaction rules matter too: who selects peers, how transactions are forwarded, and what patterns remain visible to an observer. The study shows that the interaction between Monero and Tor deserves careful attention. Its experimental results should, however, be assessed in light of the conditions under which each stage of the attack was tested.
Написать Join — чтобы участвовать в обсуждении.
wow, this is a really good write and it makes you think
good job