By Ashish Rajendra Sai

Article: “Using Artificial Intelligence To Detect Fraud On The Blockchain.

Authors: Farshad Ghassemi Toosi; Jim Buckley; Ashish Rajendra Sai; Andrew Le Gear

Affiliation: Irish Software Research Centre & Horizon Globex

Article Category: Fraud detection

Why this article? Following up on our last post, we continue our discussion of fraud detection. In this second installment, we focus on using Artificial Intelligence to identify fraud on the Blockchain. We have conducted an investigation into the potential that AI techniques that can be used in fraud detection.

Paper Overview:


Cryptocurrencies witnessed extensive attention from both academia and industry after inflation in 2017. This broad recognition has led to its adoption beyond the research domain. This increased adoption may be attributed to the overall increase in the valuation of the cryptocurrencies, which at its peak in 2017, reached a combined valuation of ~900 Billion USD. At the time of writing, the cryptocurrencies hold a combined valuation of ~200 Billion USD. This market capitalization makes cryptocurrencies very lucrative for attackers with malicious intent. A well-orchestrated attack on these cryptocurrencies may allow an attacker to attain monetary gain. The research on the security of Blockchain has identified a number of attack vectors. One such attack vector is the coordinated manipulation of the blockchain services (for example, a pump and dump scheme to artificially inflate the price of an Initial Coin Offering).

We propose using machine learning to identify the coordinated behavior of malicious entities in the blockchain network. We intend on training a neural network to identify patterns such as coordinated manipulation of the transactions to artificially inflate the hype around an Initial Coin Offering (ICO). For example, during an ICO, the company behind the ICO might try to create the impression that the offering has the potential to increase in value dramatically. They may do this by using a Hierarchical Deterministic (HD) wallet to seed new private keys, creating the impression that each purchase on the ICO comes from a different wallet/investor. This suggests that interest in the offering is of more widespread appeal than would otherwise be perceived, and may prompt other investors to purchase, resulting in an increase in the ICO’s value (Known as ’pump-and-dump’ fraud). Some work has already been undertaken, and the work of our group builds on this existing research by investigating and leveraging approaches from the field of machine learning towards the detection of pattern-based fraud on the Blockchain.


Machine learning, as an application of artificial intelligence (AI), provides a model that automatically learns from some trustable information to predict some unseen (future) data according to what was learned. The neural network is one of the several different techniques in Machine learning that comprised of a set of neurons in different layers in which the neurons of each layer interact with the neuron of the immediately next layer via some synapse. The minimum number of layers that a neural network can have is two, where the first layer accommodates the input data, and the second layer accommodates the output data. The number of neurons at the first and the last layers of each neural network indicates the dimensionality of the input and output data, respectively, e.g., a neural network with n neurons at the first layer indicates that the input data has n features (dimensions).


Our initial investigation shows how machine learning can be used to identify pattern-based frauds on the Blockchain. For our initial analysis, we investigate a Pump and Dump scheme and propose using machine learning techniques such as Linear Regression for the detection. We also propose using a strength function to categorize an ICO as fraud rather than using binary classification. Another possible use of machine learning for fraud detection is in the identification of wallets with a high degree of cohesiveness. By detecting wallets with high cohesiveness, we can identify the source of potential fraud. Conclusively, we have illustrated the advantage of machine learning techniques when practiced on the blockchain data. Given the abundance of existent machine learning techniques, we plan to do an extensive review of these techniques and the data accommodated in the Blockchain to classify possible investigation techniques and analysis targets. We then plan to coordinate these targets to the suitable techniques and appraise the effectiveness of these techniques in this novel context.

Implications for the greater blockchain community:

This article provides a fascinating insight into the world of Artificial Intelligence. The research suggests that these techniques from classical software engineering can prove to be of significant value to the blockchain world.

We ask the blockchain community if new blockchain systems should be designed to detect fraud without the need for extensive manual data analysis?

Check in for digestible insights surrounding some of the most influential research publications in the crypto/blockchain domain.

Leave a comment

Your email address will not be published. Required fields are marked *