PERSEUS exposes how masterminds manipulate meme coins in pump-and-dump schemes using AI and social networks like Telegram.

The wild world of cryptocurrency is no stranger to chaos, speculation, and, unfortunately, manipulation. One of the most infamous tactics? The age-old “pump-and-dump” scheme — reimagined for the digital age. In their 2025 paper, “PERSEUS: Tracing the Masterminds Behind Cryptocurrency Pump-and-Dump Schemes,” authors Honglin Fu, Yebo Feng, Cong Wu, and Jiahua Xu go beyond merely spotting shady activity. They zero in on something much deeper — the people pulling the strings. By introducing PERSEUS, a graph neural network (GNN)-powered system, the team takes a bold step toward identifying and exposing the actual masterminds orchestrating these digital deceptions.

The core issue tackled in this research isn’t new: pump-and-dump schemes have plagued markets for decades. What is new is the sophistication with which these schemes are executed today, often using Telegram channels and social media to mobilize masses of unsuspecting investors. The paper argues that while previous approaches have detected the events or even predicted their targets, they rarely strike at the root — the organizers themselves.

PERSEUS aims to fill that gap. It’s a detection system that combines data from online social networks (mainly Telegram) and cryptocurrency markets to build “temporal attributed graphs.” These graphs map out how information spreads, preserving who said what, when, and how it reverberated through the community. The goal: separate the real masterminds — those who originate and coordinate these schemes — from their accomplices, who merely spread the word.

The system consists of three components:

1. A real-time fetcher that scrapes data from OSNs and crypto markets.

2. A temporal attributed graph generator that builds diffusion graphs from crowd-pump messages.

3. A mastermind detector that leverages Graph Neural Networks (GAT and GraphSAGE models) to classify nodes (users) as either masterminds or accomplices.

PERSEUS was trained and tested using real-world data from 2,103 Telegram channels and 660 cryptocurrencies between 2018 and 2024, capturing over 27 million messages. It successfully identified 438 masterminds from 4,101 pump-and-dump events — and its performance, according to metrics like F1 score, accuracy, and Matthews Correlation Coefficient, beat state-of-the-art alternatives.

GraphSAGE using weighted diffusion graphs emerged as the top performer, with an impressive 0.80 precision and 0.75 F1 score. And beyond just prediction, PERSEUS gives regulators a roadmap for tracing how schemes propagate — right down to the minute and the message.

This research is impressive in scope and ambition. PERSEUS doesn’t just look for patterns; it reconstructs networks of influence in real-time. The combination of Named Entity Recognition, network graph theory, and GNNs is not only technically sound but also quite innovative. Its ability to construct time-aware, attributed diffusion graphs allows it to grasp both who spreads messages and how they spread them — which is crucial for this kind of forensic work.

Its greatest strength is that it goes after the “masterminds” — a term often thrown around but rarely quantified. By applying metrics like betweenness centrality, efficiency, and directionality of diffusion, PERSEUS can isolate those at the top of the information food chain with remarkable accuracy.

However, the research does have limitations. Mastermind detection is inherently probabilistic, and while the paper’s use of t-tests and ego-network analysis adds robustness, there’s still a layer of ambiguity. The detection system may misclassify highly active accomplices as masterminds, especially in tightly knit communities. Also, PERSEUS is currently tailored to Telegram. While this makes sense (Telegram is the epicenter of crypto coordination), broader applicability would require expansion into other platforms like Discord or Twitter (X).

Still, this paper pushes the boundaries of financial forensics, marrying AI with real-world impact in a novel and practical way.

What’s truly eye-opening is how influential a handful of masterminds can be. According to the study, just 438 individuals were linked to $3.24 trillion in trading volume increases during pump events — over eight months. That’s not just manipulation. That’s market-scale influence. Even more startling, many of these actors operate publicly, broadcasting their messages in plain sight via Telegram. The paper’s t-test analysis reveals they have higher efficiency in their personal networks, meaning most messages flow directly from them to the crowd — not through layers of intermediaries. The decentralized illusion? It’s just that — an illusion. The market is being moved by a very centralized few.

The implications here are huge — for both regulators and tech developers. For regulators, PERSEUS provides not just red flags but actual names — linked through messages, influence, and timing — that can help build legal cases. It offers a way to target enforcement at the source rather than playing whack-a-mole with accomplices or bots.

For the AI and crypto communities, this research opens doors. First, it showcases how GNNs can power real-world forensic tools in unstructured, chaotic environments. Second, it offers a replicable framework: PERSEUS could easily be adapted to stock pump groups, NFT manipulation rings, or even misinformation campaigns.

Future research might explore applying large language models (LLMs) for richer message interpretation or combining PERSEUS with blockchain analytics to track money flows, not just message flows. There’s also potential for integrating PERSEUS with decentralized platforms — giving DAOs tools to self-police.

In the fast-moving, meme-fueled world of crypto, PERSEUS shines a spotlight on the figures lurking in the shadows. By tracing the flow of messages, mapping communities, and applying GNNs to the chaos, this research gives us more than detection — it offers a plan for disruption. Masterminds aren’t mythical anymore. They’re graph nodes, identifiable, and — maybe soon — accountable.

This paper is a must-read for anyone interested in crypto regulation, AI forensics, or the battle between decentralization and manipulation. The tech is sophisticated. The implications are chilling. And the contribution? Game-changing.

Using Blockchain to Track Ransomware Threat Actors

Discover how blockchain is transforming industries on the Blockchain Insights Hub. Follow me on Twitter for real-time updates on the intersection of blockchain and cybersecurity. Subscribe now to get my exclusive report on the top blockchain security threats of 2024. Dive deeper into my blockchain insights on Medium.