Recent review studies focused on blockchain fraud detection, Graph Neural Networks (GNNs), adversarial machine learning, explainable artificial intelligence (XAI), and dynamic graph learning. While these studies have focused on various aspects such as algorithm development, model performance, or specific application domains, most studies have not extensively explored the broader challenges that stem from the dynamic nature of blockchain networks and that must be addressed to ensure the deployment of trustworthy fraud detection systems [13][14]. This study uses a Problem-based Systematic Literature Review (PBSLR) methodology to critically synthesize the techniques, security, explainability, dataset, evaluation, and regulation issues of adversarially robust and explainable GNNs for blockchain fraud detection, which is a novel approach to previous surveys. This review does not examine specific algorithms but rather summarizes general limitations, research trends, unsolved problems, and gaps in knowledge reported in studies of algorithms published in the last 5 years (2021–2026) [15]. Moreover, this paper combines five related research areas, which are researched separately and closely related to blockchain fraud detection: Graph Neural Networks, adversarial robustness, explainable artificial intelligence, dynamic graph learning, and blockchain fraud detection. The review illustrates how, when combined, these factors impact the trustworthiness, transparency, and usability of blockchain fraud detection systems [16].

The major contributions of this review are summarised as follows:

Table 1 shows that the current review articles are mostly specific to their respective research areas like Graph Neural Networks, Adversarial Robustness, Explainability, Temporal Graph Learning, or Blockchain Fraud Detection. Although these surveys give abundant information about the fields they cover, each field lacks an integrated problem-based framework considering the following areas: blockchain fraud detection, adversarial robustness, explainability, and dynamic graph learning. The focus of the present review is instead on the practical, security, explainability, and regulatory challenges, dataset issues, and evaluation challenges encountered by trustworthy GNN-based blockchain fraud detection, adopting a Problem-Based Systematic Literature Review (PBSLR) approach. Also, it puts forward an integrated research roadmap that can inform future research toward robust, explainable, scalable, and practical blockchain fraud detection systems.

The objective of this systematic literature review is to critically examine the challenges affecting the development and deployment of robust and explainable Graph Neural Networks (GNNs) for fraud detection in dynamic blockchain networks. The review focuses on five interconnected analytical dimensions: blockchain fraud characteristics, limitations of graph-based fraud detection, adversarial threats to GNN models, explainability requirements, and challenges associated with dynamic graph learning, benchmark datasets, evaluation, and deployment.

These dimensions provide the analytical structure for the literature search, study selection, data extraction, and synthesis. Rather than organizing the review around individual research questions, the literature is synthesized according to the major problems that influence the trustworthiness, effectiveness, and practical deployment of GNN-based blockchain fraud detection systems.

A comprehensive search strategy was designed to identify relevant studies published between 2021 and 2026. Multiple scientific databases were selected to ensure broad coverage of research in blockchain technology, graph machine learning, cybersecurity, artificial intelligence, and fraud detection. The selected databases included IEEE Xplore, ACM Digital Library, ScienceDirect, SpringerLink, Wiley Online Library, Scopus, Web of Science, and Google Scholar. These databases are widely recognized as authoritative sources of high-quality peer-reviewed research and have been extensively used in previous systematic reviews on blockchain and artificial intelligence [1][20]. To maximize retrieval effectiveness, Boolean operators and keyword combinations were employed. The search strings were developed based on the key concepts underlying the study, namely blockchain fraud detection, graph neural networks, adversarial robustness, explainable artificial intelligence, and dynamic graph learning.

The principal search string used was shown in Table 2 for the search combination: ("Blockchain Fraud Detection" OR "Cryptocurrency Fraud") AND ("Graph Neural Network" OR "Graph Learning") AND ("Adversarial Robustness" OR "Adversarial Attack") AND ("Explainable Artificial Intelligence" OR "Explainable GNN") AND ("Dynamic Blockchain Network" OR "Temporal Graph Learning"). Additional search combinations were executed to capture relevant studies that may have used alternative terminologies. The search process was conducted between 2025 and 2026 to ensure the inclusion of the most recent developments in the field, as shown in Figure 1 and Figure 2.

To ensure that the selected studies were relevant to the objective and analytical scope of the review, predefined inclusion and exclusion criteria were applied during the study-selection process. Establishing the criteria before full-text assessment helped to reduce selection bias and improve the transparency and reproducibility of the review process [24]. The criteria considered publication period, publication type, research relevance, methodological contribution, language, and availability of sufficient technical information, as shown in Table 3.

After the initial screening, studies that satisfied the inclusion criteria and did not meet any exclusion criterion were retained for full-text assessment. The eligible studies were subsequently subjected to quality assessment before inclusion in the final synthesis.

The study-selection process was conducted using the PRISMA framework to provide a transparent account of study identification, screening, eligibility assessment, and final inclusion. The initial database search identified 412 records from the selected academic sources. During the identification stage, 98 duplicate records were removed, leaving 314 records for title and abstract screening. Following screening, 237 records were excluded because they were outside the scope of the review, including studies that were not relevant to blockchain or graph-based fraud detection, adversarial robustness, or explainability, as well as non-peer-reviewed publications. Consequently, 77 articles proceeded to full-text eligibility assessment. The 77 full-text articles were assessed against the predefined inclusion and exclusion criteria and examined for methodological adequacy and relevance to the analytical scope of the review. At this stage, four studies were excluded: two because of insufficient methodological detail, one because it did not adequately align with the scope of the review, and one because the full text was inaccessible. This resulted in 73 studies being retained for the final systematic synthesis. The complete selection process is presented in Figure 3, illustrating the progression from the 412 initially identified records to the 73 studies included in the final review.

Following the eligibility assessment, the methodological quality and relevance of the selected studies were evaluated using a structured quality-assessment procedure as shown in Table 4. The purpose of the assessment was to ensure that the final synthesis was based on studies that provided sufficiently clear objectives, methodological information, experimental evidence, and evaluation details [20][24].

Each criterion was scored using a binary scheme, where 1 indicated that the criterion was satisfied and 0 indicated that it was not satisfied. Consequently, each study could obtain a maximum quality score of 6. The quality scores were used to assess the methodological strength of the eligible studies and to determine their suitability for inclusion in the final synthesis.

In Table 5, a minimum quality score of 4 out of 6 was adopted as the inclusion threshold. Of the 77 full-text studies subjected to quality assessment, 73 studies achieved scores between 4 and 6 and were retained for the final systematic synthesis, while four studies scoring below the threshold were excluded. As shown in Table 5, 28 studies achieved the maximum score of 6, 27 studies scored 5, and 18 studies scored 4, indicating that the majority of the included studies demonstrated satisfactory methodological quality.

A structured data-extraction process was used to collect relevant and comparable information from each study included in the final review. A standardized extraction framework was applied to ensure consistency across the selected publications. The extracted information covered bibliographic characteristics, blockchain platforms and datasets, fraud categories, GNN architectures, adversarial threats and defense mechanisms, explainability techniques, evaluation methods, major findings, limitations, and future research recommendations.

Following data extraction, the evidence was analyzed using a problem-based thematic synthesis as shown in Table 6. Rather than organizing the literature solely according to algorithms or datasets, the studies were grouped according to the recurring problems affecting trustworthy GNN-based blockchain fraud detection [25][26]. The principal themes identified included blockchain fraud characteristics, complex and dynamic graph structures, limited fraud-detection performance, adversarial attacks, insufficient adversarial robustness, explainability limitations, regulatory and compliance requirements, benchmark dataset limitations, evaluation inconsistencies, scalability, and deployment constraints [27].

The thematic synthesis was subsequently used to compare the reviewed studies, identify areas of agreement and disagreement, and determine the major unresolved research gaps. Particular attention was given to the extent to which existing approaches jointly address dynamic graph learning, adversarial robustness, and explainability, because these dimensions form the central focus of the present review in Table 7.

Blockchain technology has its origins in creating a secure, transparent, and decentralized system for digital transactions [90]. These characteristics have led to the deployment of cryptocurrencies, decentralized finance (DeFi), and blockchain services, but have also opened new avenues for blockchain-related fraud. The anonymity of blockchain transactions, along with the decentralized and borderless nature of blockchain networks, makes it hard for the typical financial monitoring systems to effectively detect suspicious activity [9]. This has made blockchain networks a popular target for a variety of financial crimes, such as cryptocurrency theft, investment fraud, money laundering, ransomware payments, illicit financial transactions, and other high-tech cybercrime scams. According to recent reports, blockchain crimes are becoming more sophisticated and extensive, with financial institutions, regulatory bodies, law enforcement agencies, and blockchain analytics firms facing serious difficulties in their attempts to combat them, resulting in billions of dollars in financial losses each year [6][12]. With the rapid development of fraudulent techniques, the complexity of pattern detection in suspicious transactions has increased, which inspires researchers to find more intelligent and adaptive fraud detection methods from the perspective of graph learning and artificial intelligence. Cryptocurrency theft and investment scams are among the most common types of blockchain fraud. They take advantage of weaknesses in the security of cryptocurrency exchanges, smart contracts, decentralized applications, or user security practices to steal digital assets or trick investors into fraudulent investment schemes [11]. These can range from fake Initial Coin Offerings (ICOs), giveaway scams, and fraudulent trading platforms to impersonation attacks and deceptive investment opportunities. As reported by [10], cryptocurrency scams in 2023 resulted in billions of dollars of stolen money, and investment fraud was one of the main causes of financial losses experienced in cryptocurrencies. Social engineering and phishing are becoming more sophisticated thanks to the increasing use of artificial intelligence. Moreover, the highly interconnected and complex transaction structures within blockchain networks make it increasingly difficult for traditional fraud detection methods to identify coordinated fraudulent activities [70][91]. Another significant obstacle in blockchain ecosystems is money laundering. Cryptocurrencies have seen a rise in their use by criminal organizations to obfuscate illegally obtained money using transaction layering, cryptocurrency mixing services, tumblers, cross-chain transfers, decentralized exchanges, and privacy-preserving cryptocurrencies. The decentralized blockchain structure enables users to send money quickly between jurisdictions without needing to traverse numerous conventional financial tracking systems, making illicit financial transactions easier [92]. Elliptic says billions of dollars’ worth of cryptocurrency transactions are laundered every year. These transactions are especially hard to detect as money laundering is done by mimicking “legitimate” financial transactions. Moreover, the advent of Decentralized Finance (DeFi) protocols has created more wash trades, making current blockchain analytics and fraud detection systems more complex.

There has also been a significant rise in Ponzi and pyramid schemes in blockchain ecosystems, notably in the decentralized finance (DeFi) sector and investment platforms using tokens as an entry point. The schemes usually offer unrealistic returns with minimal risk of investment; investment by new investors is used to pay earlier investors [93]. More recently, fraudulent investment schemes have morphed into what has been dubbed a ‘rug pull attack’, in which the developers of blockchain projects simply pull the rug from under investors after raising a lot of money. The findings have revealed that many cryptocurrency fraud losses involve fraudulent DeFi tokens and rug pull schemes [94][95]. These fraudulent projects are difficult to detect, as a large number of them appear to be similar to real investment opportunities in the early stages of development, and therefore cannot be distinguished using standard fraud detection methods [21][96]. Another significant cybersecurity issue is ransomware cryptocurrency transactions. Ransomware attacks involve the attackers encrypting victims' data and asking for a ransom, usually in cryptocurrencies, to unlock it. Cryptocurrencies are highly appealing for ransomware payments because of their decentralization, anonymity, and speed of transactions [29]. Blockchain payment channels are being increasingly used by ransomware groups to receive and distribute ransom funds, according to Europol (2024) and the United Nations Office on Drugs and Crime (UNODC, 2024). The use of intermediary wallets, cryptocurrency mixing services, and privacy-focused cryptocurrencies also makes it difficult to track ransomware payments. However, a recent study indicated that ransomware transactions tend to have unique graph structures and transaction patterns, which can enhance ransomware transaction detection models based on graphs, but accurately determining the transactions remains a challenging research problem [8][58].

In addition to fraud, blockchain networks are also used to enable widespread illicit financial flows (IFFs) such as tax evasion, dark web trading, underground market activities, sanctions evasion, and financing of terrorism. Both the Financial Action Task Force (FATF) and Europol (2024) have recently raised concerns about the use of decentralized financial (DeFi) infrastructures for transnational criminal activities. Regulatory enforcement and International financial investigations are very complex due to the cross-border and decentralized nature of blockchain transactions. Advanced graph analytics and machine learning algorithms are needed to model the multi-hop transaction networks that typically include many intermediaries, and have been shown to be applicable to the task of illicit financial flows detection [30]. The amount and sophistication of financial crimes using blockchain technology are growing, and there is a pressing need for more sophisticated, explainable, and adaptive fraud detection methods that can detect more complex fraudulent patterns in a changing blockchain environment. Overall, the literature highlights that the continuous evolution and diversification of blockchain-related fraud have identified major shortcomings in traditional fraud detection methods. The complexity, scale, and dynamic nature of blockchain transaction networks require smart graph-based learning models with the ability to capture complex transaction relationships and possess good robustness, scalability, and interpretability. The main driving force behind adversarially robust and explainable Graph Neural Networks for the purpose of accurate and trustworthy fraud detection in a dynamic blockchain environment is these persistent challenges shown in Table 8.

Figure 4 classifies significant fraud mechanisms identified in the blockchain ecosystems based on the nature of the tech fraud, such as cryptocurrency theft and scams, financial fraud schemes, fraud involving cybercrime, money laundering, ransomware-related transactions, and frauds under other categories. Based on recent blockchain crime reports and literature on the subject of fraud detection (Chainalysis, 2024; Europol, 2024; FATF, 2023; UNODC, 2024).

The blockchain transaction network is naturally modeled as a graph, as blockchain wallets, smart contracts, cryptocurrency exchanges, miners, decentralized applications, and other entities interact with each other continually through millions of transactions and blocks [30]. In the blockchain domain, the complexity of the graphs poses serious challenges for blockchain fraud detection systems [58][97][98] though Graph Neural Networks (GNNs) have shown great promise in learning from such relational data. Conventional tabular datasets are contrasted by the highly connected, high-dimensional, sparse, and heterogeneous nature of blockchain networks, which constantly change over time. This special property makes representation learning on graphs far more challenging and diminishes the impact of a lot of previous fraud detection methods. The complexity of the graph is one of the most significant issues that prevent accurate blockchain fraud detection [97][98]. The main issue is the extensive size of blockchain transaction graphs. The transaction graph on public blockchain platforms, like Bitcoin and Ethereum, expands in size and complexity with the millions of transactions and hundreds of millions of wallet addresses [27]. When the number of nodes and edges grows large, graph learning models start to experience computational bottlenecks in neighborhood aggregation, message passing, and graph convolution operations. According to [42] there are several popular GNN-based algorithms that suffer from low efficiency in large-scale blockchain networks due to memory constraints and high computational demands. In addition, fraudulent activities have long-range dependencies and complex interaction patterns across various transactions, and cannot be fully captured by shallow graph neural network architectures. Given the increasing size of blockchain ecosystems, efficient processing of large-scale transaction graphs is still a key research challenge, as shown in Figure 5.

High-dimensional node and edge features of blockchain transaction networks [32] is another significant challenge. In contrast to nodes, edge attributes are often transaction values, timestamps, gas fees, transaction directions, interaction frequencies, and transaction behavior [99] Node attributes can include wallet balances, transaction frequency, account age, smart contract interactions, token holdings, and transaction behavior. These are very rich features and useful for detecting fraudulent behavior but can also be very high-dimensional learning problems that make model development difficult. High-dimensional feature spaces lead to high computation costs, higher risk of overfitting, and lower interpretability. Several recent studies have shown that many blockchain fraud detection models are limited in their ability to select the most informative features from thousands of transaction features, as noted by [19][93]. Besides, because of the features that have redundant information, large variations in node attributes, and high correlations between features, meaningful and discriminative representations are hard to learn for nodes in the GNNs. Therefore, feature selection and dimensionality reduction still have a long way to go as open problems in the field of blockchain graph learning. The lack of data density and noise in blockchain transactions also makes it difficult to detect fraud. Many blockchain networks have a significant number of wallet addresses that have sparse neighborhood information and weak graph connectivity, as most of them are engaged in few transactions [100]. This isolation is exacerbated by the fact that bad actors often use temporary wallets and intentionally spread fraudulent actions over many wallets to evade detection. There is also a lot of noise in blockchain datasets, related to exchange wallets, smart contract interactions, inconsistent labels, missing data, and silent addresses. Proved that sparse graph structures make graph representation learning less effective since less contextual information is available for the classification of nodes [101]. Likewise, [8] have found that noisy transaction data causes classification uncertainty and results in incorrect fraud prediction. Even in the presence of noise and sparseness, however, enhancing the quality of the graph representation is still one of the core challenges for graph-based fraud detection systems.

Blockchain ecosystems are made up of highly heterogeneous entities, each with different behavioral properties, apart from their graph scale and data sparsity [9]. Individual users, cryptocurrency exchanges, decentralized finance protocols, smart contracts, liquidity pools, mining pools, token issuers, and regulatory monitoring services [21]. There are different types of transaction patterns, behavioral patterns, and risk profiles among the entities. Most of the existing graph learning methods, however, assume that blockchain graphs are structurally homogeneous, which doesn't account for the complex semantic relationships between various blockchain actors. The model Heterogeneous Graph Neural Networks (HGNNs) was introduced by [6], which models different node and edge types in a single graph framework, thus achieving better fraud detection performance. However, there are still some issues to be addressed, including scalability, consistency in representation, computational efficiency, and explainability [102][103]. Thus, modeling of the heterogeneous blockchain ecosystems in an accurate manner remains an active field of research and is still considered a key requirement for reliable fraud detection in the real-world blockchain ecosystem. In general, the literature shows that the structure of blockchain transaction networks presents significant challenges to graph learning models: very large, feature high-dimensional data, are sparse and noisy, and contain various heterogeneous entities. Each of these challenges hinders the scalability, representation capability, and detection accuracy of existing fraud detection systems based on GNNs. All of these challenges impede scalability, representational capability, and detection accuracy of the existing fraud detection systems based on GNNs. To overcome these challenges, new graph learning methods need to be created that can be scaled, are robust, and offer explainability to successfully model the complexity of the dynamic blockchain ecosystem. A comprehensive summary of the major challenges, their causes, impacts on detection performance, and corresponding research gaps is presented in Table 9.

Blockchain transaction networks are dynamic systems that continuously evolve with each passing transaction, wallet address, smart contract, and decentralized application (DApps) created [9]. In contrast to the usual graph datasets employed in machine learning, blockchain networks are dynamic, with continually changing architectures and fraud patterns, which makes fraud detection a much more difficult task [22]. Blockchain graphs grow over time, and what is learned at one moment might quickly become outdated, limiting the ability of fraud detection models [97]. Thus, temporal dependencies and reaction to the changing tactics of fraud have emerged as one of the primary challenges of graph-based blockchain fraud detection [8]. One of the primary reasons causing this problem is that constant blockchain transactions are being generated. [104] Public blockchain networks like Bitcoin and Ethereum handle millions of transactions daily, leading to the growth of transaction graphs in the network. This ongoing process necessitates models for fraud detection to adapt to node representations as they change over time in accordance with the emergence of new interactions. Most of the traditional GNN architectures, however, are static and fail to provide flexible training when there are significant changes in the graph structure, which makes them unsuitable for large-scale real-time blockchain applications. Efficiently updating graph representations is essential for maintaining accurate embeddings and ensuring the detection system performs well [29].

One more important challenge is modeling temporal dependencies. Ransomware payments, money laundering, and other fraudulent activities are often not single transactions but series of transactions over a long duration of time. Thus, learning both structural relationships and temporal interactions between blockchain entities is necessary to accurately detect these activities [29]. However, many conventional GNN models are mainly based on graph topology and fail to take temporal information into account, which restricts their ability to detect fraudulent behaviors over time. Recently, Temporal Graph Neural Networks (TGNNs) have appeared to overcome this problem, but long-term temporal dependency modeling is still an open research problem [105]. Another challenge with blockchain fraud is concept drift, which happens when fraudsters change their tactics and techniques. Over time, fraudsters constantly evolve their tactics and techniques, rendering old frauds useless and creating new fraud methods [70]. This inevitably harms the performance of machine learning models trained from past blockchain data, especially in varying settings. Showed how concept drift significantly affects the trustworthiness of blockchain fraud detection models and the importance of using adaptive learning techniques to keep up-to-date with the evolving fraud patterns in various blockchain platforms [106].

Moreover, in today's blockchain environment, timely fraud detection is essential to reduce financial losses and enable applications like cryptocurrency exchanges, decentralized finance (DeFi) platforms, and Anti-Money Laundering (AML) solutions [107]. Real-time detection is still challenging, as the learning algorithms frequently involve neighborhood aggregation, embedding updates, and message passing, which are all computationally costly; achieving a balance between detection accuracy and computational efficiency is one of the major challenges in the development of practically deployed blockchain fraud detection systems [108]. Thus, dynamic graph learning, streaming graph analytics, and online graph representation learning have become crucial research directions to develop scalable and adaptive blockchain fraud detection frameworks that can function in the real world [79]. In summary, the literature shows that current GNN-based fraud detection systems face significant challenges in light of the ongoing development of blockchain networks. Traditional static graph learning models have their limitations due to several factors, including continuous transaction generation, temporal dependencies, concept drift, and real-time processing needs. These major challenges, their impacts on fraud detection, and the corresponding research gaps are summarized in Table 10. The solution to these problems will demand dynamic, adaptive, and computationally efficient graph learning procedures that can keep up with the rapidly changing blockchain-based systems and stay accurate in fraud detection.

Although Graph Neural Networks (GNNs) have proven to be valuable tools for detecting fraud in the blockchain domain, there are still a number of limitations in their current models that hinder their accuracy, generalization, and applicability [32]. Class imbalance, classification errors, poor cross-platform adaptability, and scalability limitations are recurring problems reported by previous studies, which are still considered significant challenges in reliable fraud detection [12]. Together, these difficulties limit the utility of graph learning models in the context of dynamic blockchain settings, as criminal patterns become more intricate and more difficult to detect. The biggest issue limiting detection accuracy is the extremely skewed nature of the blockchain fraud datasets. In fact, the number of fraudulent transactions is a minor proportion of the total number of transactions on the blockchain, and thus, the class distributions are highly skewed [9]. This results in an overall high accuracy of the machine learning models and the failure to correctly classify numerous fraudulent transactions because they belong to a smaller class that is not the majority class. As a consequence, machine learning models have high overall accuracy but poor performance in detecting many fraudulent transactions because they are in a small class, which is not the most prevalent class. [21] showed that, in fact, recall and sensitivity to fraudulent activities are strongly affected if there are differences in the number of positive and negative samples; many fraudulent transactions might go unseen even when the classification performance seems to be high.

Additionally, fraud detection systems on the blockchain rely on classification errors, which reduce their trustworthiness. False positives are false alarms that occur when a legitimate user is flagged as a fraud user, interfering with financial transactions, undermining the trust of users, and increasing investigation expenses. On the other hand, false negatives enable fraudulent activities to go undetected, money to be lost, and security breaches to occur. The challenge of finding a good balance between sensitivity and specificity lies in the fact that fraudulent transaction patterns are very complex and frequently appear similar to legitimate financial transactions [15][42]. Therefore, improving false positive and false negative rates remains a key goal of blockchain fraud detection research. Poor generalization of the existing fraud detection models across different blockchain platforms is another big constraint. Most of the GNN-based models are trained and tested with data from a single blockchain ecosystem, e.g., Bitcoin or Ethereum. The transaction structures, user behaviors, consensus mechanisms, and fraud characteristics vary significantly among blockchain platforms, however [19]. Therefore, the models created for one blockchain are not likely to have similar performance on another blockchain. However, the issue of cross-chain generalization remains largely unsolved, and current blockchain fraud detection systems have not been widely used in practice, as reported by [102]. Scalability is another significant issue with the growing networks of blockchain transactions. In spite of their impressive detection accuracy in the experimental setting, many GNN models suffer from scalability issues when applied to real-world blockchain networks with millions of nodes and transactions [91]. As the size of the graph grows significantly, the memory usage, computation, training time, and inference latency also increase, making real-time deployment of the graph more difficult. Thus, it highlights the importance of scalable architectures for graph learning that can ensure high detection accuracy while efficiently handling next-generation blockchain transaction networks [8]. In conclusion, the literature shows that class imbalance, classification errors, limited cross-platform generalization, and scalability issues all contribute to the lack of reliability of current fraud detection systems based on GNNs, as summarized in Table 11. To overcome these limitations, strong graph learning models that are capable of handling highly imbalanced datasets, reducing classification errors, generalizing over heterogeneous blockchain platforms, and having computational efficiency in large blockchain environments are required.

Graph Neural Networks have substantially improved blockchain fraud detection by effectively modeling complex transaction relationships. Nevertheless, recent studies have shown that these models are highly vulnerable to adversarial attacks designed to manipulate graph structures or node features in ways that mislead the learning process [6]. Unlike conventional cyberattacks that exploit software vulnerabilities, adversarial attacks deliberately introduce carefully crafted perturbations into graph data while remaining difficult to detect. In blockchain environments, these attacks can enable fraudulent transactions, money laundering operations, and malicious wallet addresses to evade fraud detection systems, making adversarial manipulation one of the most critical security challenges confronting modern graph learning models [22]. One common attack strategy is node injection, where attackers introduce malicious wallet addresses into blockchain transaction graphs to manipulate neighborhood aggregation and graph representations. These injected nodes interact with legitimate accounts, influencing message passing and producing misleading node embeddings that reduce fraud detection accuracy [94][107] demonstrated that node injection attacks can significantly degrade node classification performance because malicious wallets are often designed to resemble legitimate users. The permissionless nature of public blockchain networks further facilitates the creation of fraudulent wallet addresses, making node injection an ongoing challenge for graph-based fraud detection systems.

Another important threat involves edge manipulation attacks, in which attackers modify graph connectivity by adding, removing, or altering transaction links. Since GNNs rely heavily on graph topology during message passing, even minor structural perturbations can substantially influence node representations and classification outcomes. In blockchain fraud detection, adversaries may create fake transaction links between legitimate and fraudulent wallets or remove suspicious connections to conceal illicit financial relationships. Demonstrated that carefully designed edge perturbations can significantly reduce fraud detection accuracy, particularly because malicious structural modifications are often difficult to distinguish from the natural evolution of blockchain transaction graphs [13][98]. Feature perturbation attacks present another significant challenge by modifying node or edge attributes without changing graph topology. Attackers may manipulate transaction values, frequencies, behavioral patterns, or account characteristics so that malicious activities resemble legitimate behavior [94]. Although these modifications are often subtle, recent studies have shown that small feature perturbations can substantially degrade GNN performance, highlighting the vulnerability of feature-based graph learning mechanisms [109]. As blockchain transaction features become increasingly complex, detecting adversarial feature manipulation becomes progressively more difficult.

Adversarial evasion attacks are performed during model deployment rather than training. Instead of modifying training data, attackers adapt transaction behaviors, distribute activities across multiple wallet addresses, and manipulate transaction sequences to avoid triggering fraud detection mechanisms [98][108]. Demonstrated that adversaries can exploit weaknesses in GNN decision boundaries to bypass detection while continuing malicious activities. The rapidly evolving nature of blockchain fraud further increases the effectiveness of evasion attacks because attackers continuously adapt their strategies in response to newly deployed detection systems. Poisoning attacks represent an additional threat during the model training phase. In these attacks, malicious or incorrectly labeled samples are deliberately inserted into training datasets to corrupt the learning process and produce unreliable fraud detection models. Within blockchain environments, attackers may inject fraudulent transaction records, fabricated wallet addresses, or manipulated graph structures into training datasets [32][110]demonstrated that poisoning attacks can substantially reduce model accuracy and increase vulnerability during deployment. Since blockchain datasets frequently contain noisy labels and imperfect transaction records, distinguishing poisoned samples from legitimate data remains an important challenge for secure graph learning.

Overall, the literature indicates that node injection, edge manipulation, feature perturbation, evasion, and poisoning attacks collectively expose fundamental vulnerabilities in existing GNN-based blockchain fraud detection systems. These attacks significantly reduce detection accuracy, compromise model reliability, and threaten the practical deployment of graph learning technologies in financial security applications. Consequently, developing adversarially robust graph neural networks has become a critical research priority for securing blockchain fraud detection systems against increasingly sophisticated attack strategies.

The Table 12 comparison shows that adversarial attacks can affect GNN-based blockchain fraud detection at different stages of the learning process. Structural attacks, such as node injection and edge manipulation, directly alter the relationships represented in the transaction graph, whereas feature perturbation modifies the information associated with existing entities or transactions. Evasion attacks primarily target the prediction stage, while poisoning attacks can influence the model during training by introducing misleading information into the learning process. Although these attacks differ in their mechanisms, they share a common objective of causing the detection model to produce unreliable predictions.

The literature in Figure 6 also indicates that the threat is particularly challenging in blockchain environments because transaction graphs are continuously updated and fraudulent actors can change their transaction behavior over time. Consequently, robustness evaluated on a static graph or under a single attack type may not adequately represent the conditions encountered in operational blockchain networks. Another important limitation is that adversarial robustness is often examined separately from explainability and dynamic graph learning. This separation makes it difficult to determine whether a model can remain accurate under manipulation while still providing reliable explanations for its decisions. Therefore, future GNN-based blockchain fraud detection frameworks need to consider structural and feature-based attacks together with temporal graph changes and explainability requirements.

Graph Neural Networks (GNNs) have demonstrated strong potential for blockchain fraud detection; however, existing architectures remain vulnerable to adversarial perturbations. Changes to graph structures or node attributes can significantly influence model predictions, reducing the reliability and trustworthiness of fraud detection systems in real-world blockchain environments [56]. This limitation is particularly important because blockchain fraud detection operates in increasingly adversarial environments in which malicious actors may deliberately modify transaction behaviors or graph relationships to evade detection. Consequently, improving adversarial robustness has become an important challenge in graph machine learning and trustworthy artificial intelligence [57].

Existing GNN architectures, including Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs), are particularly affected because their learning mechanisms depend on neighborhood aggregation and message passing. In GCNs, malicious information introduced into neighboring nodes can propagate through the graph and influence node representations, potentially resulting in incorrect classifications even when relatively small perturbations are introduced [25][58]. Although GATs use attention mechanisms to improve the representation of important neighborhood information, studies have also demonstrated that adversarial manipulation of graph neighborhoods can substantially reduce their performance under attack [7][26]. These vulnerabilities demonstrate that improving representation learning alone is insufficient to guarantee reliable fraud detection in adversarial blockchain environments.

Another important challenge is the transferability of adversarial attacks across different GNN architectures. An adversarial example designed to mislead one model may also affect other architectures, even when the attacker does not have detailed knowledge of the target model. Studies have reported that attacks developed against one architecture can transfer to other graph learning models, including GraphSAGE and GAT [35][28]. This transferability suggests that several GNN architectures share common weaknesses, making it more difficult to develop defense mechanisms that remain effective across different models. In the context of blockchain fraud detection, the problem is further complicated by the continuous evolution of fraudulent behaviors and transaction structures, which may create additional opportunities for attackers to circumvent detection mechanisms.

Developing effective defense mechanisms also introduces a trade-off between adversarial robustness and predictive performance. Techniques such as adversarial training, graph purification, robust aggregation, and certified defense can improve resistance to adversarial manipulation, but they may increase computational requirements or reduce predictive performance under non-adversarial conditions [59]. This trade-off is particularly relevant to blockchain fraud detection because practical systems must simultaneously maintain high detection performance, withstand adversarial manipulation, adapt to changing transaction structures, and operate efficiently at scale.

Overall, the literature indicates that adversarial robustness remains insufficiently addressed in existing GNN-based blockchain fraud detection approaches. Current solutions often evaluate robustness against selected attack scenarios without fully considering the combined effects of adversarial manipulation, evolving graph structures, computational constraints, and model interpretability. The major limitations related to attack coverage, graph dynamics, dataset diversity, robustness metrics, explainability, and scalability are summarized in Table 13. This creates a gap for the development of GNN-based fraud detection approaches that can maintain reliable predictive performance under both normal and adversarial conditions while remaining suitable for dynamic blockchain environments.

Graph Neural Networks (GNNs) have been found to be very good at modeling the complex network of transactions in a blockchain system, but their decision-making processes are opaque. Most GNN architectures were black-box, generating fraud predictions without offering a clear explanation of how specific transaction patterns, wallet behaviors, and/or graph structure affected the prediction. This opacity is a major hurdle for investigators, auditors, regulators, and financial institutions that rely on clear and credible proof to pursue enforcement efforts [48]. The explainability problem stems from the fact that deep graph learning models use complex message-passing and neighborhood aggregation mechanisms. These mechanisms improve the predictive performance, but they do not make it easy to know which node features, neighboring transactions, or subgraph structures are most relevant for fraud predictions [109]. This results in users being provided with only one of the two classes and with little justification, which makes it not very useful for forensic investigations or regulatory decision-making [2].

One of the other challenges is the interpretation of decision items at the node level and at the subgraph level. Often, blockchain fraud does not happen in isolation through a single wallet address, but rather through the interplay and manipulation of several wallets, smart contracts, and transaction chains. These relationships are complex and need to be explained by locating influential nodes, suspicious transaction paths, and fraudulent communities in large transaction graphs, as illustrated in Figure 7. But in many cases, the existing explanation methods fail to consistently or stably yield explanations, and subgraph interpretation is one of the least-studied topics in explainable graph learning [13][111]. These restrictions limit the confidence in automated fraud detection systems and slow down their use in high-risk financial applications. Moreover, explainability has become a key ingredient in trust, accountability, and responsible AI. Financial institutions and regulatory agencies demand high predictive accuracy for fraud detection, but they also need to be able to give clear, auditable, and human-readable explanations to help investigate fraud and ensure regulatory compliance [4]. Enhancing the explainability of GNNs has emerged as one of the most crucial research challenges to developing trustworthy blockchain fraud detection systems. Thus, the explainability of a GNN is one of the most significant research areas for trustworthy blockchain fraud detection systems. The major explainability challenges, their impacts on blockchain fraud detection, and the corresponding research gaps are summarized in Table 14.

In addition to performance, blockchain fraud detection systems need to meet ever-rising legal, ethical, and regulatory standards. With the increasing presence of AI in financial crime investigations, regulators are calling for more transparency, accountability, fairness, and auditability of automated decision-making systems [2]. The adoption of Graph Neural Networks for blockchain fraud detection is not just about prediction accuracy but also about meeting regulatory expectations. The explainability requirement from emerging AI governance frameworks is one of the key challenges. Some financial AI applications are considered to be high-risk systems under the European Union Artificial Intelligence Act, meaning that the law imposes obligations on the organization to make meaningful explanations about the decisions it can make using an automated system [112](European Commission, 2024). Many GNN models are still very opaque, making such regulatory requirements hard to meet. Likewise, the process of fraud detection should be traceable and provide sufficient evidence for any regulatory review and/or legal investigations. Financial institutions should keep clear written minutes that will allow auditors to trace a company's thinking behind fraud classifications. Deep graph learning architectures can be complex, however, which often makes decisions hard to trace back and leads to low confidence in automated systems [111]. Anti-Money Laundering (AML) regulations add to the challenges of compliance. Decentralized finance platforms, cross-chain transactions, cryptocurrency mixers, and privacy-preserving technologies are all popular features used by criminals to hide their financial operations, making it difficult for regulators to monitor them FATF, 2023Europol, 2024. In addition, ethical issues such as fairness, privacy, accountability, and possible discrimination against legitimate users arise with automated fraud detection systems [94]. The regulatory and ethical issues suggest the need for future blockchain fraud detection systems to incorporate the concepts of explainability, auditability, fairness, and privacy along with predictive functionality as summarized in Table 15.

Benchmark datasets are essential for training and evaluating GNN-based blockchain fraud detection models; however, existing datasets remain limited by class imbalance, incomplete labels, restricted fraud coverage, and limited representativeness. The Elliptic Bitcoin dataset, one of the most widely used benchmarks, contains 203,769 transaction nodes and 234,355 directed edges across 49 temporal steps, but only 4,545 transactions are labeled illicit and 42,019 licit, while most remain unlabelled [63]. Elliptic++ extends this benchmark by incorporating transaction- and address-level information, providing a richer heterogeneous graph representation, although the large proportion of unknown labels remains a challenge for supervised learning [64]. More recently, Elliptic2 introduced approximately 49 million node clusters, 196 million transaction edges, and about 122,000 labeled subgraphs for money-laundering analysis; however, its emphasis on anti-money laundering limits its coverage of other fraud categories such as phishing, ransomware, Ponzi schemes, and smart-contract fraud [65]. Similar limitations occur in Ethereum fraud research, where datasets are frequently constructed independently from blockchain transaction records, resulting in differences in observation periods, labeling procedures, feature extraction, and graph construction that make cross-study comparison difficult [66]. Although some blockchain datasets provide timestamps or predefined temporal steps, they generally represent fixed historical observations rather than continuously evolving transaction streams, limiting realistic evaluation of concept drift and continual learning [32]. Furthermore, few benchmark datasets provide standardized adversarial settings for evaluating node injection, edge manipulation, feature perturbation, poisoning, and evasion attacks [31][59]. Consequently, the literature demonstrates a need for larger and more diverse multi-platform benchmarks containing reliable fraud labels, temporal transaction evolution, multiple fraud categories, and standardized protocols for evaluating predictive performance, adversarial robustness, and explainability. The characteristics, strengths, and major limitations of representative blockchain fraud detection datasets are summarized in Table 16.

The evaluation of trustworthy Graph Neural Networks is getting increasingly complicated, as the classic performance measures (such as accuracy, precision, recall, F-1 score, ROC-AUC) reflect merely a part of the model performance [98]. Several other criteria like adversarial robustness, explainability, scalability, fairness, and reproductivity of a blockchain fraud detection system have to be taken into account to assess their merit. But, at present, there is no universal acceptance of a framework for the evaluation of the different approaches, which makes them hard to compare in objective terms [111]. There is a lack of standard measures of robustness. Previous research uses various metrics, such as attack success rate, robustness accuracy, perturbation budgets, certified robustness, and robustness degradation, which limits the ability to make cross-study comparisons [83]. Such limitations also apply to the evaluation of explainability, in which measures like fidelity, sparsity, stability, completeness, and human interpretability are used in varying ways across different studies [2][3]. The additional challenge of comparative evaluation is that the benchmark datasets vary, as do the way they are generated, plotted, and manipulated to create graphs, the feature engineering approaches, the methods used to generate attacks, and the experimental protocols used [113]. Furthermore, many published works do not provide access to source code or implementation details and/or benchmark datasets, resulting in large reproducibility hurdles and independent verification of the reported results [19]. These evaluation limitations need to be addressed to develop the trustworthy standards of benchmarking and fast-track progress towards a trustworthy blockchain fraud detection system. The major evaluation challenges, their impacts, and the corresponding research gaps are summarized in Table 17.

Numerous blockchain security studies have established that detecting fraud within a blockchain has emerged from a classical classification problem to a multi-disciplinary research task that includes graph representation learning, cybersecurity, explainable artificial intelligence, adversarial machine learning, and financial regulation [14][114]. Although GNN-based fraud detection has made many strides, previous research shows that there are various recurring issues that still are not sufficiently addressed. In a technical sense, blockchain transaction networks can be characterized as large-scale graphs with heterogeneous entities, high-dimensional node and edge attributes, sparse interactions, and dynamically evolving topologies. In actual blockchain networks, these factors add significant complexity to calculations and affect their expandability of the current persistent security challenges identified as adversarial attacks were also found in the review. Surprisingly, although high predictive accuracy is desirable, existing GNN models are not robust against malicious manipulation, even being vulnerable to node injection, edge manipulation, feature perturbation, poisoning attacks, and evasion attacks, as can be seen in [4]. This is because blockchain fraudsters are constantly tweaking their methods to bypass detection systems, and ensuring that blockchain systems become more adversarial-robust by adapting their methods to evade detection is necessary for practical deployment. Explainable graph learning models are another important aspect that we found lacking in existing models. While effective classification performance of many GNN designs, the black-box nature of the process hinders transparency and limits their use in financial settings where regulatory compliance, auditability, and trust are critical [10][11]. The existing Explainability Methods are still very experimental and have failed to deliver consistent explanations for real-world financial investigations. Another key limitation, in agreement with the review, was the use of benchmark datasets, which is still problematic in blockchain fraud detection research. Previously collected datasets tend to be static, very unequal in distribution, and have low counts of confirmed fraudulent transactions. Furthermore, there are very few benchmark datasets available to use for evaluating adversarial robustness and/or dynamic graph learning, which makes it challenging to compare the performance of graph learning algorithms objectively [38][47]. All of these results suggest that research has come a long way, making great strides in enhancing the accuracy of fraud detection but has received comparatively less focus on considering robustness, explainability, dynamic adaptability, and practical deployment requirements. As a result, numerous methods are still not viable with real-life blockchain networks.

The reviewed literature published between 2021 and 2026 shows a clear shift from conventional machine-learning approaches based on manually engineered transaction features toward graph-based learning methods that can capture relationships among blockchain entities. GNNs have become increasingly important because they enable end-to-end learning from transaction graphs and can represent complex fraud patterns that are difficult to identify using traditional classifiers [16]. More recent studies increasingly focus on dynamic and temporal graph learning, adversarial robustness, and explainability. This indicates a broader transition from accuracy-oriented fraud detection toward trustworthy AI systems that are expected to combine predictive performance with robustness, transparency, scalability, and regulatory compliance. The trend therefore suggests that future blockchain fraud detection research will increasingly favor integrated frameworks rather than isolated improvements in individual model components.

The reviewed studies indicate important trade-offs among detection accuracy, adversarial robustness, explainability, and computational efficiency. More complex GNN architectures may improve predictive performance, but they often increase training cost, inference time, and deployment complexity. Similarly, adversarial defence methods such as adversarial training and graph purification can improve resistance to manipulation, but may reduce clean-data performance or require additional computational resources [69][70]. Explainability introduces another challenge because explanations must be sufficiently detailed to support investigation and regulatory requirements without oversimplifying complex graph relationships. Therefore, future research should treat accuracy, robustness, explainability, and efficiency as interconnected objectives rather than independent optimization problems.

The review shows that different areas of GNN-based blockchain fraud detection have reached different levels of maturity. Graph representation learning and GNN-based fraud classification are comparatively well established, with numerous studies demonstrating their effectiveness in modelling complex transaction relationships [21]. Dynamic and temporal graph learning has also advanced considerably, although continual learning, concept drift, and streaming analysis remain open challenges. In contrast, adversarial robustness and explainability are less mature. Robust GNN methods are often evaluated under limited attack settings and rarely on realistic blockchain datasets, while explainability approaches remain largely focused on local explanations that may not fully support financial investigations [11]. Overall, the literature suggests that the most important unresolved direction is the integration of dynamic learning, adversarial robustness, and explainability within a unified and practically deployable GNN framework.

A key finding of this review is that adversarial robustness and explainability should not be considered independent research objectives. Instead, they're considered as supplementary traits of reliable AI systems. Robustness allows to ensure that fraud detection models are immune to adversarial manipulation while explainability allows investigators, financial institutions, and regulatory agencies to understand and verify model decisions [10]. If an explanation is not robust, it could describe a false prediction; if a prediction is robust but not explainable, either within or beyond the high-stakes financial domain, it may be hard to justify. Many recent studies have indicated that combining robustness and explainability in a single graph learning framework could lead to better system reliability and gain stakeholder’s trust. But there are very few studies which explicitly explore this relationship, creating a great opportunity for future study.

Significant strides have been made, however, there are key issues that remain to be solved. These encompass the construction of scalable GNN architectures, that could process large dynamic graphs, with provable robustness against adaptive adversarial attacks, benchmark datasets to ground truth the robustness of these models, unified robust evaluation protocols, and explainability frameworks. Moreover, the level of demand for transparency, auditability, fairness, and compliance keeps rising, which imposes even more demands on AI-driven fraud detection solutions [12]. This does not only need industry collaboration and partnerships between researchers in graph machine learning, cyber security, blockchain technology, XAI and financial regulation, it will also need a deeper interdisciplinary integration of these domains. Based on the results of this review, the key research recommendations are the need for further research into integrated, accurate and trustworthy GNN frameworks that can meet regulatory compliance, scalability, explainability, robustness and accuracy requirements simultaneously and deployed easily in practice.

Existing GNN-based fraud detection approaches primarily emphasize predictive performance, while robustness, explainability, fairness, uncertainty, and regulatory requirements are often considered separately. Future research should develop trustworthy graph-learning frameworks that jointly optimize these requirements rather than treating them as independent components. Such frameworks should integrate robust learning, interpretable decision-making, uncertainty estimation, fairness-aware modeling, and adaptive learning to support reliable deployment in real-world blockchain environments.

The review shows that existing GNNs remain vulnerable to structural and feature-based adversarial manipulation, while many defence mechanisms are evaluated against limited attack scenarios. Future studies should investigate adaptive adversarial training, robust message-passing mechanisms, graph purification, certified robustness, and self-supervised defence strategies. More importantly, these methods should be evaluated against multiple and adaptive attacks using realistic blockchain transaction graphs rather than relying primarily on static or synthetic benchmark settings.

Explainability remains insufficiently developed for practical blockchain fraud investigation. Existing approaches often provide local feature- or subgraph-level explanations that may not adequately communicate why a transaction, wallet, or group of entities has been classified as suspicious. Future explainable GNNs should provide interpretable transaction paths, influential entities, temporal fraud patterns, and prediction confidence while incorporating domain knowledge from financial investigators. Human-centred evaluation is also necessary to determine whether explanations are understandable and useful for investigation, auditing, and regulatory decision-making.

Blockchain transaction networks continuously evolve as new transactions, wallets, smart contracts, and decentralized applications emerge. Consequently, models trained on historical graph snapshots may lose effectiveness as transaction patterns and fraud strategies change. Future research should investigate temporal GNNs, continual and incremental learning, concept-drift adaptation, online graph learning, and streaming graph analytics. These approaches should enable fraud detection systems to adapt to emerging transaction behaviors without requiring complete model retraining.

The lack of representative and standardized benchmark datasets remains a major limitation in comparing blockchain fraud detection approaches. Future benchmark datasets should incorporate multiple blockchain platforms, diverse fraud categories, temporal transaction evolution, reliable labels, and realistic adversarial scenarios. Standardized evaluation protocols are also required to assess predictive performance, adversarial robustness, explanation quality, scalability, computational efficiency, and reproducibility under comparable experimental conditions. Such benchmarks would improve the reliability and comparability of future studies.

Future research should move beyond offline transaction classification towards intelligent systems capable of continuously monitoring blockchain networks and detecting suspicious activities in real time. Such systems could integrate dynamic GNNs, adversarial defence, explainability, streaming graph processing, and decision-support mechanisms within a unified architecture. Important capabilities include early fraud detection, continual adaptation, interpretable alerts, scalable processing, and support for regulatory auditing. Achieving this objective will require collaboration across graph machine learning, blockchain security, cybersecurity, explainable AI, distributed systems, and financial regulation.

Based on the research gaps identified in this review, the development of GNN-based blockchain fraud detection can be viewed as a gradual transition from performance-oriented models towards trustworthy, adaptive, and deployable fraud intelligence systems. In the short term (2026–2028), research should prioritize robust GNN architectures, improved graph explainability, dynamic benchmark datasets, and standardized robustness and explainability metrics. In the medium term (2028–2031), attention should shift towards integrating robustness, explainability, and continual learning within unified frameworks, while improving scalability and cross-blockchain generalization. In the long term (beyond 2031), these developments could support autonomous real-time fraud intelligence systems capable of continuous adaptation, adversarial resilience, interpretable decision-making, and human–AI collaborative investigation. A consolidated summary of these research gaps, proposed directions, and expected outcomes is presented in Table 18.

The proposed research directions emphasize that future progress should not be measured solely by improvements in fraud classification accuracy. The development of trustworthy blockchain fraud detection requires the coordinated advancement of dynamic learning, adversarial robustness, explainability, scalability, and standardized evaluation. Addressing these gaps will provide a stronger foundation for reliable GNN-based fraud detection in evolving blockchain environments.