Early Rug Pull Detection on Solana

arXiv:2611.13342 · cs.CR · Submitted 2026-11-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "Early Rug Pull Detection on Solana".

Elias: Here's the gist of the paper. Early rug-pull detection on Solana is investigated by comparing graph neural networks against non-graph models to determine if token transfer structure adds predictive value at minute-scale decision points The results show that the early signal is real, but no evidence suggests the transfer graph makes detection earlier than a model without it Pruning redundant reasoning tokens The study evaluates whether token-transfer structure adds predictive value over non-graph models at minute-scale decision points The researchers compare Graph Neural Networks (GNNs) with matched non-graph models that receive the same token-level and graph-summary features on identical samples, including a fixed panel of tokens observed in all three windows They address two research questions: RQ1 Across windows: does a GNN at q1 outperform XGBoost at q1, q2, or q3, i.e., detect rug pulls earlier?

Nadia: and RQ2 Within a window: does a GNN outperform a non-graph model that has observed the token for the same time?

Elias: The results show that the early signal is real, but no evidence that the transfer graph makes detection earlier How the training signal is built The operational rug-pull label requires three conditions to hold jointly: (1) Liquidity drop, (2) Price collapse, and (3) Persistent depletion over a nominal 24-hour horizon after creation The three observation windows are the 25th, 50th, and 75th percentiles of the Solana rug-pull lifespan distribution of Chen et al.

7: , measured on 76,469 rug-pull tokens from the first half of 2025: wq1 = 750 s, wq2 = 2,127 s, and wq3 = 5,585 s Token-level features are computed from Birdeye trades in [t0, tq) and cutoff liquidity snapshots (Table 1) Transfer graphs are constructed where each retained transfer is one directed edge with relative time phi = (t - t0)/wq in [0, 1) Where the gains come from The structural differences between rug-pull and non-rug transfer graphs already differ clearly after 12.5 minutes, most strongly in receiver-side centralization Rug-pull graphs involve more endpoints (median 145 vs.

Title and authors: Nadia: Let's start with who wrote this: Saba Heydaridoost, Kasra Kashani, Mohammad Mowlavi Sorond, and Behnam Bahrak from the Tehran Institute for Advanced Studies in Tehran. They're the researchers behind "Early Rug Pull Detection on Solana."

Elias: It’s interesting seeing research coming out from that kind of setup. The title itself sets up a very specific problem: catching those rug pulls early on Solana, which is where things move incredibly fast.

Priya: When you look at the authors, it's good to see a mix of backgrounds. We have applied security researchers like Nadia and cryptographers like Elias who understand the technical side, and privacy folks who focus on what the data actually represents.

Nadia: Exactly. The fact that you have people focused on both exploitation defense and understanding private measurement really helps frame this kind of study about token legitimacy.

Elias: I wonder what assumptions they made about the system they were testing? Because when you’re building a detection model, the assumptions you start with can totally skew what you find.

Priya: That's a fair point. I want to know if their label definition was robust enough to catch real scams without accidentally flagging normal trading activity as suspicious.

Nadia: The paper lays out exactly how they defined the operational label, which is pretty important because that's where you determine what counts as a rug pull in their experiment.

Elias: They used three conditions to hold together for a rug pull label: liquidity drop, price collapse, and persistent depletion over a nominal twenty-four hour horizon after creation.

Priya: That sounds like a pretty standard set of criteria, but I'm curious about the specific windows they chose for those observations.

Nadia: They used the twenty-fifth, fifty percent, and seventy-five percentiles of the Solana rug pull lifespan distribution to set their observation windows: two thousand five hundred seconds, two thousand one hundred twenty seven seconds, and five thousand five hundred eighty five seconds.

The paper's summary: Nadia: So the paper is essentially comparing Graph Neural Networks against non-graph models to see if the token transfer structure provides any edge in detecting rug pulls at minute-scale decision points.

Elias: They set up a comparison where both models get the exact same features, including the token-level data and graph summaries, on identical samples across different observation windows.

Priya: The main research questions they tackle are whether a Graph Neural Network outperforms an XGBoost model at any of those specific time cutoffs, and whether a GNN performs better within a single window compared to a non-graph model observing the same token over time.

Nadia: Their findings conclude that the early signal is real, but they didn't find any evidence that the transfer graph itself makes detection happen earlier than models that don't use it.

Elias: That’s a key distinction they make: there’s an indication of an early event, but the structural information from the graph doesn't seem to give them a head start in prediction.

Priya: So, even though the signal is real, if I'm building a system right now, I shouldn't rely on the transfer graph being superior for that initial minute-scale warning.

Nadia: That’s right. It suggests that for now, focusing on those other features might be more efficient than adding the complexity of a full transfer graph analysis at that exact moment.

Elias: They also noted some structural differences between rug pull graphs and non-rug transfer graphs already appearing after twelve point five minutes, particularly concerning receiver-side centralization.

Priya: That's interesting because it suggests the *pattern* of the network changes relatively quickly, even if the overall predictive advantage doesn't materialize immediately for their models.

The paper's improvements: Nadia: The paper points out a few practical things for platforms that list new Solana tokens. First, model scores computed from the very first minutes of activity can help prioritize tokens for review, but we have to stress that these aren't reliable probabilities or definitive judgments.

Elias: They also found that data availability is tied to the outcome; tokens with complete cases and those that generate transfer graphs actually have a higher rug pull rate.

Priya: That’s a practical observation about data quality, which is something I think every privacy researcher needs to hear—the quality of what you collect directly affects what you can actually learn.

Nadia: Then there's the point about temporal shifts: the observed shift in activity over time motivates platforms to keep monitoring performance and reassessing how they train and calibrate their models as the token population changes.

Elias: And they also issue a caution that any claims of improved detection from graph-based scoring should be backed up by paired comparisons, multiple seeds, and test sets large enough to actually see those meaningful gains.

Priya: That’s a necessary grounding for any claim. It means we can't just trust the output of one specific model variant without rigorous testing against other setups.

Nadia: Exactly. The paper is nudging platforms toward better monitoring and more cautious interpretation of these early warning scores rather than treating them as certain predictions.

Conclusion: Elias: So to wrap this up, the main question was whether token transfer graphs enable earlier rug-pull prediction for newly launched Solana tokens than matched non-graph models.

Nadia: The answer they found is that the rug pull and non-rug graphs already differ structurally within twelve point five minutes, which is something important to know about how these scams behave on the network.

Priya: From my angle, I think it’s a good finding that even if the GNN doesn't win yet, understanding that structural difference after just twelve point five minutes gives us insight into the token's initial behavior.

Elias: And from a cryptographic side, it means we have to be careful about what assumptions we make about the data structure when trying to build these predictive tools for Web3.

Nadia: It’s a reminder that the early signal is real, but it doesn't automatically mean the tool that finds it earliest is the one you should rely on blindly.

Priya: I think looking at those structural differences in receiver-side centralization after twelve point five minutes gives us a concrete pattern to watch for when we're analyzing token activity.

Elias: So, this paper on "Early Rug Pull Detection on Solana" shows that while the graph structure is interesting and reveals early patterns, it doesn't currently provide an advantage in getting the absolute earliest detection compared to simpler models.

Saba Heydaridoost, Kasra Kashani, Mohammad Mowlavi Sorond, Behnam Bahrak

Tehran Institute for Advanced Studies

cs.CR

Submitted: 2026-11-09

Updated: 2026-11-09

Code: https://github.com/Saba-Heydaridoost/rugpull_detector

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 67/100

The gist: Early rug-pull detection on Solana is investigated by comparing graph neural networks against non-graph models to determine if token transfer structure adds predictive value at minute-scale decision…

Key concepts

Graph Neural Networks (GNNs)
These are machine learning models specifically designed to analyze relationships between entities, like tokens and their transfers. They look at the structure of how tokens move from one to another, treating the token network as a graph. This helps in understanding complex patterns in on-chain activity.
Rug-Pull Label
A specific definition for when a rug pull has occurred. It requires three things to happen together: a sudden drop in liquidity, a sharp price collapse, and sustained depletion of the token over at least 24 hours after its creation. This strict criteria defines the event being studied.
Transfer Graphs
These are visual representations of token movement. Each transfer is an edge connecting two tokens, and the relative time between transfers forms a value on that edge. Analyzing these graphs helps identify structural differences between normal trading activity and potentially fraudulent schemes.
Receiver-Side Centralization
This refers to how many incoming transfers a specific token receives from others. The study found that rug-pull graphs show different patterns here compared to non-rug graphs, suggesting that the way tokens are being collected or distributed is a key structural difference early on.

Terminology

Summary

Early rug-pull detection on Solana is investigated by comparing graph neural networks against non-graph models to determine if token transfer structure adds predictive value at minute-scale decision points The results show that the early signal is real, but no evidence suggests the transfer graph makes detection earlier than a model without it

Pruning redundant reasoning tokens

The study evaluates whether token-transfer structure adds predictive value over non-graph models at minute-scale decision points The researchers compare Graph Neural Networks (GNNs) with matched non-graph models that receive the same token-level and graph-summary features on identical samples, including a fixed panel of tokens observed in all three windows They address two research questions: RQ1 Across windows: does a GNN at q1 outperform XGBoost at q1, q2, or q3, i.e., detect rug pulls earlier? p. 0

Where the paper says it

We address two research questions: RQ1 Across windows: does a GNN at 𝑞1 outperform XGBoost at 𝑞1, 𝑞2, or 𝑞3, i. , detect rug pulls earlier?

and RQ2 Within a window: does a GNN outperform a non-graph model that has observed the token for the same time? p. 0
Where the paper says it

We pre-specified this comparison as our primary analysis before evaluating on the test data. RQ2 Within a window: does a GNN outperform a non-graph model that has observed the token for the same time?

The results show that the early signal is real, but no evidence that the transfer graph makes detection earlier

How the training signal is built

The operational rug-pull label requires three conditions to hold jointly: (1) Liquidity drop, (2) Price collapse, and (3) Persistent depletion over a nominal 24-hour horizon after creation The three observation windows are the 25th, 50th, and 75th percentiles of the Solana rug-pull lifespan distribution of Chen et al. [7], measured on 76,469 rug-pull tokens from the first half of 2025: wq1 = 750 s, wq2 = 2,127 s, and wq3 = 5,585 s Token-level features are computed from Birdeye trades in [t0, tq) and cutoff liquidity snapshots (Table 1) Transfer graphs are constructed where each retained transfer is one directed edge with relative time phi = (t - t0)/wq in [0, 1)

Where the gains come from

The structural differences between rug-pull and non-rug transfer graphs already differ clearly after 12.5 minutes, most strongly in receiver-side centralization Rug-pull graphs involve more endpoints (median 145 vs. p. 0

Where the paper says it

Rug-pull and non-rug transfer graphs already differ clearly after 12. 5 minutes, most strongly in receiver-side centralization.

58.5 in q1) but are sparser The largest differences concern receiver-side centralization: the maximum in-degree centrality and maximum PageRank are clearly higher in non-rug graphs (r = −0. p. 4
Where the paper says it

295] GAT -. 212] The largest differences concern receiver-side centralization: the maximum in-degree centrality and maximum PageRank are clearly higher in non-rug graphs (𝑟 = −0.

42 for both in q1) However, none of the three GNNs at q1 is significantly better than XGBoost at any of the three cutoffs (RQ1), and XGBoost itself does not improve with longer observation

Implications for Web3 platforms

The findings suggest four practical points for platforms that list new Solana tokens First, model scores computed from the first minutes of activity can help prioritize tokens for review, but should not be treated as reliable probabilities or definitive judgments Second, data availability is associated with the outcome: complete cases and graph-bearing tokens have higher rug-pull rates Third, the observed temporal shift motivates ongoing performance monitoring and reassessment of training and calibration as the token population changes Fourth, claims of improved detection from graph-based scoring should be supported by paired comparisons, multiple seeds, appropriate multiple-testing correction, and test sets large enough to detect practically meaningful gains

Limitations

The label covers rug pulls that drain liquidity within 24 hours; slower rug pulls and schemes without a liquidity drain are out of scope The tokens come from one SolRPDS file, mostly from 2024, and the transfer graphs are partial The comparisons rely on a single chronological split with test sets of 136–159 tokens Class weighting was tuned for the graph models but not for all non-graph models; re-tuning XGBoost with class weighting did not change the primary conclusion One candidate feature, liquidity at creation, was dropped after the split had been fixed because it was missing for 94–97% of samples The model variants of Section 5.5 were designed after the primary results were known; their comparisons are exploratory and not corrected for multiple testing

Conclusion

The study asked whether transfer graphs enable earlier rug-pull prediction for newly launched Solana tokens than matched non-graph models Rug-pull and non-rug graphs already differ structurally within 12.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Early signal prioritization: The system can compute model scores from first minutes of activity to help prioritize tokens for review, but these scores should not be treated as reliable probabilities or definitive judgments.

  2. Structural pattern recognition: The system can identify that rug-pull and non-rug transfer graphs already differ clearly after 12.5 minutes, most strongly in receiver-side centralization, informing users about the structural differences in token activity early on.

  3. Model selection guidance: The system can use a pre-specified comparison of graph and non-graph models to determine whether a GNN at q1 outperform XGBoost at q1, q2, or q3, i.e., detect rug pulls earlier, providing an initial assessment of the transfer graph's predictive value.

  4. Monitoring data quality: The system can monitor input coverage alongside model scores because data availability is associated with the outcome, suggesting platforms should watch for changes in token population and missingness as they occur over time.

  5. Temporal reassessment: The system can be designed to trigger ongoing performance monitoring and reassessment of training and calibration because the observed temporal shift motivates ongoing performance monitoring and reassessment of training and calibration as the token population changes.

Abstract

Web3 platforms that help users discover and trade newly launched crypto tokens, including launchpads, decentralized-exchange (DEX) interfaces, wallets, and token-analytics sites, must assess tokens before their legitimacy is clear. On Solana, where token creation is inexpensive and tokens use standard programs rather than custom per-token smart contracts, rug pulls can unfold within minutes. Effective warnings must therefore rely on limited histories. Although graph neural networks (GNNs) have improved rug-pull detection on EVM chains, it remains unclear whether transfer structure adds predictive value over non-graph models at minute-scale decision points, and how that value changes as history accumulates. We study 2,700 tokens sampled from SolRPDS, collecting one-minute market histories from Birdeye and token transfers from Solana RPC. We define a rug pull operationally by requiring a liquidity drain, price collapse, and persistent liquidity depletion over a nominal 24-hour period after token creation. All features use only information available at or before each prediction cutoff. We evaluate observation windows of 12.5, 35.5, and 93 minutes, selected in advance from the published lifespan distribution of Solana rug pulls, and represent each token as a temporal transfer graph. Within the first 12.5 minutes, rug-pull and non-rug graphs already differ in size, density, reciprocity, and receiver-side centralization. We compare GNNs against matched non-graph models using the same token-level and graph-summary features on identical samples. A fixed panel of tokens observed in all three windows further isolates the effect of longer observation from changes in token population. Our findings inform the design of early-warning mechanisms for newly launched Solana tokens in Web3 interfaces.

Sources

Related papers