Diva.exchange is building a value-exchange system rooted in user privacy, anonymity, and security. That is why it relies on I2P (Invisible Internet Project) as its overlay network – a message-based anonymous network resistant to censorship and restrictions.

But how reliable is this protection when artificial intelligence is deployed against it?

The anonymity of any network layer requires regular stress-testing. As part of a collaboration with the Lucerne University of Applied Sciences and Arts (HSLU), researcher Patrick Grossenbacher examined the potential of machine learning for deanonymizing I2P users (“Machine Learning Approaches to Deanonymizing I2P Network Traffic”). The research results were presented at the CRITIS conference on September 28–30, 2026.

Below is a breakdown of the methodology, key findings, and implications of the study for the diva.exchange architecture.

Threats to Anonymous Networks

Attacks on anonymous networks fall into two categories: active (introducing malicious nodes, compromised floodfill routers) and passive.

The study focuses exclusively on passive analysis. In this scenario, the adversary does not interfere with network operations, but merely records encrypted network traffic at the internet service provider (ISP) or backbone link level.

We set out to determine whether an observer can identify the source and destination IP addresses of a communication session (outbound gateway and inbound endpoint) without cracking the encryption, relying solely on timestamps and encrypted packet sizes.

Put simply, can artificial intelligence establish the fact that information is being exchanged between two nodes in the network?

I2P Architecture and the Risk Zone

Unlike Tor, I2P operates on a packet-switched data transmission model using unidirectional tunnels. Protection is built on garlic encryption and noise-based NTCP2/SSU2 protocols. Data is transmitted inside fixed-size encrypted blocks (up to 1 KB), which conceals its internal structure.

However, I2P has a specific characteristic: the network does not use mandatory variable latency or persistent cover traffic by default.

How Artificial Intelligence Searched for Connections Between I2P Network Users

To collect data, a test network of 16 i2pd routers was deployed in a Kubernetes cluster. All network packets were captured with nanosecond precision (tcpdump --nano). Machine learning algorithms then attempted to “connect the dots” and identify the actual sender and recipient.

The analysis took place in two steps:

1. Searching for similar signals. The algorithm examined millions of transmitted data packets and attempted to guess which ones belonged to the same route.
2. Time verification. If two nodes continuously exchange information at the same rhythm for several minutes, the probability of error decreases.

Key Findings of the Study

1. Full control gives algorithms an advantage

With 100% traffic visibility, deanonymization is possible. In a laboratory setting, intercepting absolutely all network packets allowed the algorithm to reconstruct up to 99% of sender-recipient links with an accuracy of up to 93% under a strictly configured confidence threshold.

In other words, if a hypothetical observer (such as a major telecom provider) sees absolutely all traffic in the network, algorithms are indeed capable of linking the sender and recipient with high precision. Under ideal laboratory conditions, the model was able to locate the majority of participants.

2. If only a portion of the network is visible, the defense holds up

As soon as the observer loses sight of even a portion of transmissions (for example, failing to control all nodes or missing some data), the algorithm’s accuracy drops sharply. Finding connections becomes nearly impossible. Losing even part of the packet chain makes it difficult to reconstruct the full path.

3. The more users in the network, the higher the security

The experiment was conducted in a test network of 16 nodes. In real life, the I2P network consists of tens of thousands of participants. In such a “thick fog,” millions of simultaneously transmitted messages merge into a unified noise, making it significantly harder to isolate the required route.

Conclusions and Prospects for diva.exchange

The study demonstrates that time-correlation analysis poses a theoretical challenge under total traffic visibility, while simultaneously confirming that I2P’s architecture holds up resiliently against passive attacks in realistic, distributed environments.

For diva.exchange and the broader ecosystem, these findings highlight three key priorities:

• Enforcing Modern Transport Standards: Consistently relying on SSU2 and NTCP2. Both protocols leverage the Noise Protocol Framework and natively feature cryptographic padding mechanisms to obfuscate packet sizes and message boundaries against passive observation;
• Network Scaling as Primary Defense: The research confirms that algorithm precision collapses rapidly as network density increases. Expanding the number of active nodes and participant volume is the most effective natural countermeasure, as higher background entropy renders route isolation practically infeasible;
• Application-Level Architecture: Rather than altering the underlying low-latency network layer, diva.exchange can decouple application workflows. Implementing asynchronous transaction bundling and event-driven request scheduling naturally eliminates predictable communication rhythms at the source.

Research Materials

All research materials – including the labeled network packet dataset, traffic cleanup scripts, and model training code – are published under the MIT license in open access:

• Dataset: i2p-dataset-2026
• Model code: i2p-deanonymization-ml

THIS IS DIVA.EXCHANGE

The non-profit association diva.exchange, Switzerland, uses a barrier-free and collaborative approach to create free banking technology for everyone. Open-source technology ensures the privacy of all participants in the financial system of the future. The blockchain-based system is fully distributed. Everyone can participate in diva.exchange.

Diva.exchange is committed to the belief that only commercially free technology can reliably protect user privacy.

Collaboration with the scientific community plays an important role in the development of diva.exchange. The results of diva.exchange research are constantly being validated by academic institutions and publicly presented at specialized conferences.

LEARN MORE ABOUT OUR WORK

All technical information is available at: https://github.com/diva-exchange/

I2P beginner’s guide and installation guide:https://www.diva.exchange/en/privacy/introduction-to-i2p-your-own-internet-secure-private-and-free/

All videos are here: https://odysee.com/@diva.exchange:d/

Introduction to I2P: https://en.wikipedia.org/wiki/I2P

Testnet of diva.exchange: https://testnet.diva.exchange

CONTACT US

Twitter: https://twitter.com/@DigitalValueX

Mastodon: https://social.diva.exchange/@social

If you still have questions you can always find us on Telegram: https://t.me/diva_exchange_chat_de (in English, German, or Russian)