ABC: A Cryptocurrency-Focused Threat Modeling Framework

by   Ghada Almashaqbeh, et al.
NYU college
Columbia University

Cryptocurrencies are an emerging economic force, but there are concerns about their security. This is due, in part, to complex collusion cases and new threat vectors, that could be missed by conventional security assessment strategies. To address these issues, we propose ABC, an Asset-Based Cryptocurrency-focused threat modeling framework capable of identifying such risks. ABC's key innovation is the use of collusion matrices. A collusion matrix forces a threat model to cover a large space of threat cases while simultaneously manages this process to prevent it from being overly complex. We demonstrate that ABC is effective by presenting real-world use cases and by conducting a user study. The user study showed that around 71 identify financial security threats, as compared to only 13 who used the popular framework STRIDE.



There are no comments yet.


page 9


Information Security Risks Assessment: A Case Study

Owing to recorded incidents of Information technology inclined organisat...

A Framework for Data-Driven Physical Security and Insider Threat Detection

This paper presents PS0, an ontological framework and a methodology for ...

A Graph Based Framework for Malicious Insider Threat Detection

While most security projects have focused on fending off attacks coming ...

Multi-Stage Threat Modelling and Security Monitoring in 5GCN

The fifth generation of mobile networks (5G) promises a range of new cap...

Towards automation of threat modeling based on a semantic model of attack patterns and weaknesses

This works considers challenges of building and usage a formal knowledge...

Security for Emerging Miniaturized Wireless Biomedical Devices: Threat Modeling with Application to Case Studies

The landscape of miniaturized wireless biomedical devices (MWBDs) is rap...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Cryptocurrencies and blockchain technologies are an emerging economic force. Since Bitcoin emerged in 2009 [33], the number of these “digital currencies” has grown into the thousands in 2018, with a total market capital exceeding 130 billion [9]. Although early systems focused only on providing a virtual currency exchange medium [33, 12], nowadays there is increasing interest in providing other types of distributed services, such as computation outsourcing [10] and file storage [17]. These newer applications suggest increased adoption of cryptocurrency-based systems in the coming years.

Yet, despite the many advantages they offer — decentralization, transparency, and lowered service costs — there is still a big gap between the promise of cryptocurrencies and their performance in practice. A major stumbling block is the perception that these systems are not secure, and the large number of security breaches announced in the past few years give credence to these doubts [15, 11, 5, 6, 7, 13, 16]. Therefore, a better understanding of the security of cryptocurrencies is needed in order to ensure their safe deployment in emerging applications.

The best practice for designing a secure system requires a threat modeling step to investigate potential security risks. Such a model can guide system designers in deploying the proper countermeasures, and assessing the security level of a system. Although threat modeling has been thoroughly studied in the literature, existing paradigms primarily target software applications [28] or distributed systems that have a small number of participant types [32]. These threat modeling techniques were frankly not designed to be scalable for a set of diverse, mutually distrustful parties as found in cryptocurrencies. Such systems, especially these providing distributed services, consist of parties that play different roles (miners, clients, and different types of servers), and an attacker may control any subset of these parties. Adding to the complexity, the attacker may seek to target any role in the system and may launch a diverse array of attacks with different intended outcomes. The complexity of reasoning about and managing threat cases becomes unwieldy as these sets grow.

To address this issue, we propose ABC, an Asset-Based Cryptocurrency-focused threat modeling framework. ABC introduces a novel technique, called a collusion matrix, that allows users to investigate the full threat space and manage its complexity. A collusion matrix is a comprehensive investigation and threat enumeration tool that directly addresses collusion by accounting for all possible sets of attacker and target parties. ABC helps in reducing the combinatorial growth of these cases by ruling out irrelevant ones and merging threat cases that have the same effect. Such explicit consideration of attacker collusion is particularly important in permissionless cryptocurrencies that allow anyone to join.

ABC’s models are also tailored to better consider the threat domain of cryptocurrencies. This is done by introducing new threat categories that account for the financial motivations of attackers and the new asset types, i.e., critical system components, these systems introduce. ABC identifies these categories by listing the assets in a system, such as the blockchain, and outlining the secure behavior of each asset. Then, the threat categories are defined as any violation of the security requirements for these assets. This approach produces a series of system-specific threat classes as opposed to an a priori-fixed list of generalized ones. Another feature of ABC is acknowledging that financial incentives and economic analysis can play major roles in other steps in the design process. These tools can be used in risk assessment and in mitigating some types of attacks that cannot be neutralized cryptographically.

To demonstrate the framework’s effectiveness, we conducted a user study and prepared use cases. The study compared the performance of subjects in building threat models using ABC and the popular framework STRIDE. Among the obtained results, we found that around 71 of those who applied ABC were able to identify financial threats in a cryptocurrency system, as compared to less than 13 of those applying STRIDE. In addition, while none of the STRIDE session participants spotted collusion between attackers, around 46 of those who used ABC identified these scenarios. For the use cases, we applied ABC to three real-world systems, including Bitcoin, Filecoin, and CacheCash. These cases attest to the usefulness of ABC’s tools where they integrated well into CacheCash’s design phase, and they revealed several missing threat scenarios in the public design of Filecoin. These findings confirm the potential of ABC as an effective tool for assessing and improving the security of cryptocurrencies.

2 Related Work

To orient readers to the current state-of-the-art in threat modeling, we summarize here some prior work done in this area. We also highlight relevant works on threat identification and security analysis in cryptocurrencies.

Threat modeling frameworks. The STRIDE framework is one of the most popular strategies in this field [28]. STRIDE is an acronym of the threat categories it covers, namely, Spoofing, Tampering, Repudiation, Information disclosure, Denial of service (DoS), and Elevation of privilege. This framework is a multistep procedure that involves understanding the software application functionality, capturing its operation flow using data flow diagrams (DFDs), mapping the components of these DFDs to the threat categories mentioned previously, and employing threat tree patterns to derive concrete threat cases. Though several solutions have extended STRIDE to accommodate more complex systems [32], and cover other security requirements, e.g., privacy [22], its model does not fit cryptocurrencies. Another study [39] has borne out this premise, where the authors extended STRIDE’s threat categories to handle Bitcoin-like community currencies. However, their approach targets only community fund operation, where more modifications would be needed to handle other components of the currency exchange medium.

Other paradigms have pursued slightly different approaches. KAOS [38] is a goal-oriented requirements engineering framework that has been extended to cover security. ANOA [20] is a generic framework to define and analyze anonymity of communication networks. And the framework presented in [27] targets the secure design of data routing protocols. Finally, other works build specialized threat models for specific classes of distributed systems, e.g., storage systems [25], rather than introducing a framework. These works indicate that different types of systems have different requirements when performing threat modeling. This reinforces the idea that emerging systems, such as cryptocurrencies, need specialized threat modeling tools.

Security analysis of cryptocurrencies. Most of the work done so far in this category can be divided into two classes. The first formalizes the security properties of consensus protocols and blockchains [34], while the second discusses specific security attacks on cryptocurrencies. For example, in a series of studies on Bitcoin, Bonneau et al. [21] present several security threats, Androulaki et al. [18] evaluate its anonymity property, and Gervais et al. [24] study how tampering with the network links affects participants’ view of the blockchain. Other works dealing with different types of cryptocurrencies include Luu et al. [30] who focus on Ethereum smart contract security, and Sanchez et al. [31] who analyze the security of Ripple.

While these attack descriptions are very useful, they only outline specific threat scenarios for a given system. Our goal, however, is to develop a framework that allows reasoning about the full set of potential attacks facing a variety of cryptocurrency types.

3 Stepping through the ABC Framework

Having highlighted the need for a cryptocurrency-specific threat modeling framework, we now present the ABC model. As a systematized approach, applying ABC starts by understanding the functionality of the cryptocurrency system under design with a focus on its asset types and the financial motivations of the participants (Section 3.1). This is followed by identifying the impactful threat categories and mapping them to the system assets (Section 3.2). After that, ABC directs system designers to extract concrete attack scenarios using a new tool called a collusion matrix, which helps in exploring and analyzing the full threat space (Section 3.3). Lastly, ABC acknowledges that financial incentives affect other design steps including risk assessment and threat mitigation (Section 3.4).

To make the discussion easier to follow, we illustrate the ABC process by describing its application to the following simplified system, which was inspired by Golem [10]:


is a cryptocurrency that provides a distributed computation outsourcing service. Parties with excessive CPU power may join the system as servers to perform computations on demand for others. Clients submit computation jobs to servers, wait for the results and proofs of correctness, and then pay these servers cryptocurrency tokens. The mining process in CompuCoin is tied to the amount of service provided to the system. That is, the probability of selecting a server to mine the next block on the blockchain is proportional to the amount of computation it has performed during a specific period.

The full threat model of CompuCoin is available online [14]. Several excerpts from this model are embedded in the discussion of ABC steps that follows.

3.1 System Model Characterization

Understanding the system is an essential step in the threat modeling process. A misleading or incomplete system description can lead a designer to overlook serious threats and/or incorporate irrelevant ones. Therefore, an accurate system model must outline the use scenarios of the system, the assumptions on which it relies, and any dependencies on external services. In addition, the system model must be aware of all participant roles, and any possible motivation they might have to attack the system. For the latter, evaluators need to consider how the financial interests of these participants shape their behaviors.

Moreover, a system model must define the critical components that need to be protected from attackers. These components represent the assets that would compromise the whole system if attacked. To capture the features of the system, ABC identifies these assets based on its functionality. In detail, ABC divides the system into modules, and labels the valuable components of each module as assets. These assets could be concrete or abstract resources [32]. For example, the blockchain and the currency can be considered concrete assets, while preserving user privacy would be an abstract asset.

Finally, a system model includes graphic illustrations of its work flow. For distributed systems, it is useful to draw network models [32], in which system modules are represented by graphs showing all participants and assets, and the interactions between them. These graphs are helpful when enumerating the concrete threat scenarios as we will see in Section 3.3.

Running example application. Figure 1 illustrates how this step would look in CompuCoin. It shows the various components of the system model, in addition to a network model of the computation outsourcing service. It should be noted that the asset list in this figure is not exhaustive, and is limited by the rather brief description provided for CompuCoin. Moreover, these assets could be divided differently. For instance, one may merge service payments and the currency together. However, we believe that a fine-grained division provides a more comprehensive threat treatment.

Figure 1: System model characterization of CompuCoin.

3.2 Threat Category Identification

After understanding the system model, the next step is to identify the broad threat categories that must be investigated. For each system component, system analysts outline all threat classes that may apply. Here ABC steps away from conventional practice of using an a priori-fixed list, and instead uses an adaptive approach inspired from requirements engineering [22] that defines threats as violations of system security goals. Given that assets are the target of security breaches, ABC defines these threat classes as violations of asset security requirements. This allows deriving cryptocurrency-specific threat categories because ABC identifies the assets in a way that aligns with the functionality of these systems.

Asset Security Threat Category
Service Service corruption (provide corrupted service for clients).
Denial of service (make the service unavailable to legitimate users).
Information disclosure (service content/related data are public).
Repudiation (the server can deny a service it delivered).
Service Service slacking (a server collects payments without performing all the promised work).
payments Service theft (a client obtains correct service for a lower payment than the agreed upon amount).
Blockchain Inconsistency (honest miners hold copies of the blockchain that may differ beyond the unconfirmed blocks).
Invalid blocks adoption (the blockchain contains invalid blocks that does not follow the system specifications).
Biased mining (a miner pretends to expend the needed resources for mining to be elected to extend the blockchain).
Transactions Repudiation (an attacker denies issuing transactions).
Tampering (an attacker manipulates the transactions in the system).
Deanonymization (an attacker exploits transaction linkability and violates users’ anonymity).
Currency Currency theft (an attacker steals currency from others in the system).
Network Denial of service (interrupt the operation of the underlying network).
Table 1: CompuCoin threat categories.

Accordingly, ABC examines each asset in the system and applies the following procedure to identify its threat classes:

  • Define what constitutes secure behavior for the asset, and use that knowledge to derive its security requirements. These requirements include all conditions that, if met, would render the asset secure. For example, CompuCoin’s servers provide a computation outsourcing service and collect payments in return. We may consider the service payment asset secure if: a) servers are rewarded properly for their work, and b) that they earned the payments they collected.

  • Define the threat categories of an asset as violations of its security requirements. Tying this to the above example, the service payment asset would have the following threat classes: service slacking, where a server collects payments without performing all the promised work, and service theft, where a client obtains service for a lower payment than the agreed upon amount.

The previous steps are highly dependent on how system analysts define the security properties of an asset, especially if there is no agreed-upon definition in the literature. For example, several works studied the security of the blockchain [21, 34]. Yet, there is no unified security notion for the service asset because each type may have different requirements.

Running example application. Applying this step to CompuCoin produced the threat categories listed in Table 1 (the detailed process can be found in Appendix G). We found this table useful when building threat models for all the use cases found in Section 5, where we mapped the listed categories to the assets in each system. In this process, we found that some threat types were not applicable due to the absence of some assets. Notably, Bitcoin’s only assets are the ones related to the currency exchange medium. On the other hand, some systems required replicating some of these categories among all instances of an asset, e.g., in Filecoin all service asset threats were replicated for the two service types this system provides, namely, file storage and retrieval. This shows how the system characteristics affect the threat category identification step in ABC.

3.3 Threat Scenario Enumeration and Reduction

Once the threat categories have been identified, the next step is to enumerate concrete attack scenarios under each threat type. It is important in this step to be as comprehensive as possible by considering all potential attackers, target parties, and the set of actions attackers may follow to achieve their goals. This also involves considering collusion between several attackers who may cooperate on attacking the system.

Detecting collusion is particularly important in cryptocurrencies. This is because the presence of monetary incentives may motivate attackers to collude in more ways than traditional distributed systems. The popular centralization problem caused by mining pools attests to this fact, where these pools can collude and perform devastating attacks. Even miners may collude by accepting, or rejecting, updates on the network protocol which leads to hard forks in the system. ABC can detect these, and other, collusion cases at early stages of the system design.

To achieve this, ABC introduces collusion matrices that instruct analysts to enumerate all collusion cases, and reason about the feasibility of all threats scenarios in the system. A collusion matrix is two-dimensional, with the rows representing potential attackers and the columns representing the target parties. For the rows we list all participant roles in the system, both individually and in every possible combination. We also add a category called “external” that represents all entities outside the system. The same is done for the columns, with the exception that “external” is excluded. By definition, an external party is not part of the system, and hence, can not be a target. Each cell in these matrices represents a potential threat case to be investigated.

An example of a collusion matrix for the service theft threat in CompuCoin is shown in Figure 2. The dashed ellipse in the accompanying network model encloses a service session, which is an interaction between a server and a client. Any entry with multiple parties on the attacker side in this matrix indicates collusion. Note that a participant label may represent slightly different roles depending on where it is placed. For example, in Figure 2 the label “server” on the target side corresponds to a single server (i.e., Bob) since a service session involves only one server. However, the label “server” on the attacker side represents all servers in the system, including Bob. Hence, the cell in grey shade in Figure 2 does not suggest that a server colludes to attack itself, but instead, it represents the case where other servers collude with Alice against Bob.

Figure 2: Collusion matrix of service theft threat in CompuCoin.

For each threat category mapped to the assets in the system, a separate collusion matrix is created and analyzed as follows:

1) Enumeration: In this step, system analysts examine each cell and enumerate all strategies that attackers can use against the target parties, and documenting the process. It is useful to consider the network model of the system components as they show the interactions between the participants and the system assets.

2) Reduction: While examining each cell, system analysts reduce the number of threat cases by:

  • Eliminating cells representing scenarios that will not produce a threat to the system. This consists of crossing out the eliminated cells and documenting the rationale for elimination. For example, in Figure 2, the cells that have the client as a target are irrelevant to the service theft threat. This is because a client does not provide a service to others. Other cases can also be crossed out if they are neutralized by system assumptions or by early design choices. For example, requiring all transactions to be signed by their originators rules out transaction repudiation and tampering attacks.

  • Merging together scenarios (and the corresponding cells) that have the same effect, or those that do not become stronger with collusion. For example, in Figure 2, the grey shaded cell in which Alice is colluding with other servers to avoid paying Bob is reduced to the case that Alice is a sole attacker. This is because only Alice pays for the service she receives from Bob, while other servers are not part of the protocol.

3) Documentation: System analysts should document all threat scenarios that remain after the reduction step. That is, each documented case should outline the attack description, the target asset(s), the attacker(s), the flow of actions, all preconditions that make the attack feasible, and the reasons behind merges and deletions (if any).

The overall number of matrices and the size of each matrix depend on the system parameters, such as number of participant roles and assets. The above reduction steps eliminates a substantial number of cells in a documented, principled way, which saves time and effort.

Figure 3: Analyzing service theft matrix in CompuCoin.

Running example application. The CompuCoin threat model has 11 collusion matrices [14]. We present one of them here; the service theft threat collusion matrix as illustrated in Figure 3. As shown, 21 cells can be reduced to just 2 threat scenarios (merged and ruled-out cells are displayed in pink and black shades, respectively). In this matrix, ten cases have been ruled out. This includes all cells under the column with the “client” header, for the reasons explained previously, and the first three cells under the column with the “server” header. This is because “external” and/or “server” cannot be attackers because they do not ask/pay for the service111One may say that an external may join the system as a client to perform the attack. This case is covered under the client role in the matrix..

Ten other merged cases are shown in Figure 3. This includes all cells under the column with the “client and server” header, which are reduced to attacking only servers. This is again because clients do not serve others. The rest of the merges cover the last three cells in the column with the “Server” header. In these cells, a client is colluding with an external entity and/or other servers to make the target server lose payments. Such collusion will not make a client stronger (these parties can drop/withhold payments, however, this is covered under DoS threat). Hence, all these cells are reduced to the case of a solo client attacker.

3.4 Risk Assessment and Threat Mitigation

The outcome of the threat modeling process, i.e., the documented list of impactful threat cases, can provide the designers with a guiding map to secure the system. During this process, it is useful to prioritize threats based on the amount of damage they can cause. This falls under the purview of risk management, a separate task from threat modeling, carried out using frameworks like DREAD [28].

ABC integrates with risk management by leveraging existing techniques for threat mitigation. For example, many threat vectors can be addressed using rational financial incentives that are often called detect-and-punish mechanisms. That is, when a cheating incident is detected, the miners punish the attacker financially. These approaches can use a game theoretic approach [37] to set the design parameters in a way that makes cheating unprofitable. By modeling interactions between the players as an economic game, the financial gain of all player strategies can be computed. Then, the parameters are configured to make honest behaviors more profitable than cheating. The same procedure can be used to quantify the damage these financial threats may cause. In other words, a threat that could give the attacker a big payoff should be prioritized over a threat that yields minimal profits. This reinforces the idea that cryptocurrencies require an expanded model for exploring risks and countering them.

Running example application. To illustrate this step in CompuCoin, we consider the distilled threat scenarios found in Figure 3. Both threats can be neutralized financially by designing proper techniques to make the client lock the payments in an escrow, along with a penalty deposit. The client loses the latter if he should cheat, perhaps by issuing invalid payments that carry his signature. The penalty deposit amount can be computed as the maximum payoff a client may obtain by cheating. This makes cheating unprofitable, and hence, unappealing to rational clients.

4 Evaluation

To evaluate the effectiveness of ABC, we set up an empirical experiment that compares how it performs against STRIDE [28], a widely used threat modeling framework. We chose STRIDE for this comparison because it is a popular example of the type of a model a system designer will turn to in the absence of a cryptocurrency-specific framework [39]. The experiment took the form of a user study in which participants were asked to build threat models for a simple cryptocurrency system using one of these two frameworks. As our primary goal was to test whether financial incentives and collusion could influence the type of threats discovered, our evaluation focuses on answering the following questions:

  1. Does a threat modeling framework affect how subjects characterize a system model?

  2. Do the threat categories of each framework influence the broad threat classes identified by the subjects?

  3. Do participants build more accurate threat models when using ABC than when using STRIDE?

  4. Do participants find the ABC/STRIDE method easy to use in completing the study?

In what follows, we discuss the study methodology and some of the insights drawn from the findings.

4.1 Methodology

We recruited 53 participants, primarily masters students in systems security programs. We used five subjects as a pilot group to test and refine our materials. The remaining 48 participants were divided randomly into two groups of 24, one of which built the threat model with STRIDE, while the other used ABC.

Each testing session spanned three hours and was divided into two parts: a group tutorial and individual completion of threat models. The group tutorial started with a 20 minute overview of cryptocurrencies, followed by a one-hour training in the framework to apply. The participants were then given a 25 minute break to reduce any fatigue effects. The session resumed with a 15 minute overview of ArchiveCoin, the system for which subjects will build a threat model. ArchiveCoin is a simplified Filecoin [17]-inspired cryptocurrency system that focuses mainly on the service and its rewards in order to fit the study session period.

The individual completion of threat models spanned the remaining hour of the study session. Given that the allocated session period was short, we asked the subjects to look into just one threat category in Step 3, namely, the service theft of file retrieval. This category was not used in the clarifying examples of the tutorials to avoid biasing the results. Participants performed Steps 1 and 2 (system model characterization and threat category identification), and then submitted their answers. Only at this point were they given the materials for Step 3, in which they were asked to eliciting threat scenarios for service theft of file retrieval. This was done so that participants who missed this threat when answering Step 2 could not alter their responses. At the end, the participants were asked to fill out a short questionnaire in which they rated how easy or hard it was to apply the threat modeling framework they employed. Our study instrument and all supporting materials are available online [14].

4.2 Findings

In this section, we discuss the main results produced by the study.

4.2.1 Effect on the System Model Characterization Step.

In the first step of each threat modeling framework, the subjects were asked to characterize the system model by defining its modules, its assets, and the participant roles, in addition to drawing either a network model of the system, in case of ABC, or a data flow diagram (DFD), in case of STRIDE. To quantify the influence of the framework on this step, we compute the subject scores in the first step based on reference threat models we built for ArchiveCoin222We built two reference models, one using STRIDE and one using ABC to evaluate the responses of each framework session. Nonetheless, both models produced the same list of elicited threat cases in the last step.. We report these scores after normalization, meaning that we divide them by the maximum score value one may obtain when answering everything correctly.

Figure 4: Subject scores for Step 1. Diamonds indicate the mean.

The results for Step 1 are found in Figure 4333This figure is a box plot [8]

, which displays the distribution of the data points by showing the maximum and minimum (the whiskers above and below the box), the median (horizontal line inside the box), and the data points that span the first to third quartiles (the box itself). In case most of these points are very close this box is suppressed into a line.

. As shown, ABC scored higher than STRIDE, with total average values of 0.92 and 0.69, respectively. Analyzing the responses for the sub-steps in Step 1 revealed several interesting observations. The first one is related to identifying the financial assets and modules in the system. As the figure shows, several subjects who applied STRIDE did not identify the payments (or currency) as an asset. Instead, their focus was on the user files stored in the system. Similarly, most of them did not identify the payment process as a system module, and focused only on file storage and retrieval processes. On the other hand, most of the subjects in the ABC session identified these financial related assets and modules, as reflected in the lower average scores of STRIDE participants reported in Figure 4, and the frequency results found in Figure 5. These results indicate that employing conventional threat modeling frameworks, instead of ones that are customized for monetary-incentivized systems, could lead evaluators to neglect the financial aspects of the system. This, in turn, could cause evaluators to miss important threat cases, and thus, leave the system vulnerable to attacks.

Figure 5: Identifying payment related modules and assets.

The second observation is related to how subjects defined the participant roles in the system. As shown in Figure 4, STRIDE achieved an average score of 1 in this category as compared to only 0.89 for ABC. All STRIDE session subjects defined the participant roles correctly, which in that model, included only clients and servers. For ABC, we mentioned in its tutorial that an “external” entity must be considered among the participant roles. Yet, not all the subjects in the ABC session recorded the “external” role in their responses. This points to an important observation. Evaluators may only consider insider attackers because they interact with the system, and forget that external entities could be also motivated to, and capable of, an attack. Considering external attackers affects not only the eliciting of concrete threat scenarios in Step 3, but also the identification of the threat categories in Step 2. Therefore, more emphasis need to be placed on this role early on in the threat modeling process.

Lastly, the third observation is related to the framework influence on how subjects represented the system modules graphically. Figure 4 shows that the average scores for the network model/DFD sub-step were found to be 0.98 and 0.44 for ABC and STRIDE, respectively. STRIDE session subjects struggled to draw a DFD for ArchiveCoin because such a representation is more suitable for software applications than distributed systems. On the other hand, ABC’s use of network models made this task easier for its session subjects, and almost all of them sketched diagrams correctly. As mentioned previously, this graphic representation helps in eliciting the concrete threat scenarios in the system (Step 3), and hence, inaccurate diagrams may affect the outcome of this process.

4.2.2 Effect on the Threat Identification Step.

In Step 2 of both frameworks, the subjects were asked to define the broad threat categories to be investigated. As part of the study material, participants who applied STRIDE were given its threat category list, along with the component mapping table found in the STRIDE user guide. On the other hand, participants who applied ABC were given the list found in Table 1 (covering only the service and service reward assets). The subjects in both groups defined the categories to be considered for ArchiveCoin by mapping these lists either to the system assets (in case of ABC), or to the DFD components (in case of STRIDE). The reference models we built indicated that the mapping outcome for both frameworks would include the following threat classes: service corruption, DoS, information disclosure, service slacking and service theft for both service types that ArchiveCoin provides (file storage and retrieval).

Figure 6: ABC and STRIDE scores for Step 2. Diamonds indicate the mean.

Based on the scores for the threat identification step found in Figure 6 the cryptocurrency-tailored categories of ABC made it easier for the study participants to identify the threat categories in question as compared to STRIDE, despite the subjects having little experience with cryptocurrency-based systems. The average score for ABC subjects is around 0.51, compared to 0.29 for STRIDE (note these scores are normalized as mentioned before). The generalized categories used by STRIDE fit software applications well, but they do not suit monetary-incentivized distributed systems. System analysts using these generalized categories would need to expend more time and effort in order to identify the more specific threat classes we are interested in.

To provide more insights= about the effectiveness of the threat categories of each framework, we analyzed the number of subjects who identified each threat category as depicted in Figure 7. ABC is ahead of STRIDE for all financial-related threats, i.e., service slacking and theft, as well as the service corruption threat. For the service theft of file retrieval threat, which is the category that we asked the participants to investigate in Step 3, only three participants in the STRIDE session spotted this threat, while 17 subjects in the ABC session did so, or around 13 and 71, respectively. For service theft of file storage and slacking of file retrieval, none of STRIDE participants spotted these threats, and only one participant spotted service slacking of file storage. On the other hand, 67%, 17%, and 50% of ABC participants identified these threats, respectively. This, again, shows that the ABC categories guided the subjects toward service and payment related threats in a better way than the general threat classes included by STRIDE.

Figure 7: Frequency of threat category identification.

For the rest of the threat classes, we found that STRIDE’s subjects are ahead of ABC’s session participants for both DoS and information disclosure threats. As shown in Figure 7, around and of STRIDE subjects identified these categories, respectively, while around and of ABC’s subjects did so. Although we do not have a precise justification for this outcome, we think that this is due to the fact that these threats are thoroughly explained in the threat category table of STRIDE, which includes attack examples as well. This allowed STRIDE participants to pay more attention to these threat types. Hence, we believe that the ABC tutorial needs to stress these threats and explain them in greater depth.

4.2.3 Threat Model Accuracy.

To quantify accuracy, we compute the recall and precision values for the concrete threat scenarios found by each subject as compared to the reference threat models we built for ArchiveCoin. The recall is computed as , and precision is computed as , where a true positive is a correctly identified threat, false negative is an undetected threat, and false positive is an incorrectly defined threat. The recall (precision) indicates how many valid (invalid) threats a subject defined. Both quantities take values between 0 and 1.

Based on the results of Step 3 in each framework (i.e., eliciting concrete threat scenarios), we found that participants who applied ABC produced a larger number of valid threat cases than STRIDE session subjects, with average recall values of 0.48 and 0.4, respectively. At the same time, participants using ABC identified a lower number of irrelevant cases than those who applied STRIDE. The former scored an average precision value of 0.57, as opposed to 0.48 for the latter444One participant in ABC session has 0 false negative and 0 true positive, we excluded him/her when we computed the average.. We believe that this is due to several factors. First, ABC directed participants to consider the financial aspects of the system, which affected the elicited threat scenarios. Second, the use of the collusion matrices helped ABC participants to reason about the threat space in an organized way that reduced random speculations, as opposed to the STRIDE threat tree patterns that work well when applied to software applications. Third, ABC’s collusion matrices guided participants to spot threat cases caused by collusion, as opposed to STRIDE’s tree patterns that focus only on solo attackers. The results show that none of the subjects in the STRIDE session identified a possible collusion case between a client and servers, while 11 subjects in the ABC session identified this collusion case555Most of them, however, did not provide a clear description of the attack scenario. Hence, these incomplete descriptions were not counted as correct threats when grading Step 3.. This confirms the importance of considering collusion when investigating threat cases, and shows the usefulness of ABC matrices in handling this task.

Figure 8: Total normalized scores (diamonds indicate the mean).

All these factors affected the overall correctness of the threat models built by the subjects. As shown in Figure 8, the ABC session scored an average of 64 as compared to 50 for STRIDE. This is expected based on the reported results for the modeling steps, where ABC scores were ahead of those of STRIDE.

4.2.4 Framework Ease-of-use.

As mentioned previously, at the end of each session, we asked participants to report on how easy they found applying the framework in question. Ease-of-use was measured on a Likert scale in which 1 indicates the lowest value and 5 indicates the highest value. The average values are 3.9 for ABC and 3.8 for STRIDE666Three participants did not complete the questionnaire in the STRIDE session.. This result becomes somewhat more significant when we point out that the participants already had some exposure to STRIDE and its threat tree patterns. Even though it is a new framework that introduced several new concepts to the participants, ABC still achieved a comparable ease of use level. This suggests participants were able to grasp its concepts through just a single hour of training, and therefore ABC shows potential as a usable method for threat modeling.

4.3 Threats to Validity

We acknowledge that a few limitations must be kept in mind when considering the study results. Empirical studies of threat modeling usually span a longer time frame, often on the orders of months, e.g., [35]. However, the fact that we were able to glean several important observations suggest that there may be lessons to be learned from short focused studies. In addition, we feel the design of our study might serve as a guide for determining promising areas for extended research before a large commitment of time and resources is made. Another constraining factor could be the age and experience level of our subjects. Different responses might have been obtained if we tested system security experts. However, we believe that the inexperience of our participants match the cryptocurrency space well, which attracted users and researchers from various fields, even those from outside systems security. As such, our results give indications on how they might perform when investigating the security of cryptocurrencies.

5 Experiences

To demonstrate how ABC would function when applied to complex real-world systems, we developed use cases in which we built threat models for three cryptocurrencies. Each of these cryptocurrencies represents different stages in a system design lifetime. Bitcoin [33] is a well-established system, Filecoin [17] is under development and close to being launched, and CacheCash is still in its early development stages. The analyses for Bitcoin and Filecoin stopped before the risk management/mitigation phase. However, CacheCash’s analysis involves also the risk mitigation step as we later describe.

5.1 Bitcoin Analysis (Steps 1-3)

Bitcoin is by far the most valuable cryptocurrency with a capital market share of around 66 billion as of March 2019 [9]. As shown in Table 2, the Bitcoin threat model has significantly fewer collusion matrices and threat cases than other systems777The full threat model of Bitcoin is available online [14]. This is because it provides only a currency exchange service, which reduces the number of assets. Furthermore, it involves only two types of participants, miners and clients, which reduces the size of the collusion matrices. These factors, in addition to our familiarity with Bitcoin design details, contributed in reducing the completion time of Bitcoin’s threat model as shown in the table. Moreover, at the time we were working on this model, we had already completed the design of ABC. This suggests that deep understanding of the system model, and the availability of suitable tools impact not only the accuracy of the results, but also the time and effort expended in the threat modeling process.

Aspect Bitcoin Filecoin CacheCash
ABC steps covered Steps 1 - 3 Steps 1 - 3 Steps 1 - 4
Completion time (hr) 10 47 Not tracked
No. of collusion matrices 5 14 9
Total threat cases 105 882 525
Distilled threat scenarios 10 35 22
Table 2: Threat model comparison.

We drew two main observations about the threat model we generated for Bitcoin. First, all the known threats to Bitcoin, such as double spending, Eclipse attacks [26], Goldfinger attack [29], and delaying blocks and transaction delivery [24], were mapped to the collusion matrices produced by ABC. Second, collusion between participants can play a major role in Bitcoin’s security. That is, several threats are neutralized by the assumption that at least 50 of the mining power is honest. Yet, mining pools have been formed to concentrate mining power. At the time of this writing, around 95 of the mining power is in the hands of just 10 mining pools [2]. If the managers of these pools decide to collude, they could break the honest majority barrier and take the system down. In fact, serious security attacks can be performed with less amount of mining power. Sompolinsky et al. [36] attest that a selfish mining attack, or blocks withholding, in which an attacker controls even less than 30 of the mining power, would be able to undermine the fairness of the mining reward distribution.

Furthermore, miner collusion may take a different form represented by rejecting specific updates on the network protocol. As cryptocurrencies are still in the development stage, features and updates are still being added to their protocols. These updates can create soft and hard forks in the network [19]. In this form of collusion, a subset of the miners do not agree to adopt the new version of the protocol. This causes a split in the network, i.e., inconsistency in the blockchain view, and may extend to spinning out a new version of the currency. This happened in Bitcoin, where there two new cryptocurrencies split from its network including Bitcoin Cash [3] and Bitcoin Gold [4].

Usually, the 50% threat, or miners’ collusion in general, is argued about informally using incentive compatibility. This idea asserts that rational miners are more interested in keeping the system running to preserve the value of their rewards. However, this claim is hard to verify and remains as an open question [21]. In addition, this assumption might be valid when all parties belong to the same system. Yet, miners could be working in several systems and it could be the case that destroying one to strengthen the other would be more profitable. Nevertheless, such observations highlight two key points. First, it indicates the importance of validating all the security assumptions a system makes in its design. And second, it points to the need for rational economic incentives to address some types of security threats that cannot be addressed by using only cryptographic approaches. The design of ABC accounts for such observations as mentioned earlier in the paper.

5.2 Filecoin Analysis (Steps 1-3)

Filecoin [17] is a cryptocurrency-based system that provides distributed data storage and retrieval services. Any party may join the system as a storage or retrieval miner to offer service to others. Filecoin operates distributed retrieval and storage markets where clients and miners can submit storage/retrieval bids and offers. Once these offers are matched, the service-payment exchange process, in which clients pay the miners in the Filecoin currency in exchange for receiving correct service, may start. The mining process in Filecoin is tied to the storage service miners put in the system. Recently, the Filecoin team raised around 250 million through an ICO (initial coin offering) [1] in preparation for an official launch.

Filecoin is a more complicated system than Bitcoin as it provides two types of services on top of the currency exchange medium. In addition, its protocol involves three participant roles: clients, retrieval miners, and storage miners, with the latter filling the traditional roles of miners in maintaining the blockchain. This complexity is reflected in the number of collusion matrices and threat cases produced as shown in Table 2. Moreover, all threat categories that target the service asset were replicated for each service type, which contributed to the large size of the threat model. These factors affected the completion time to build the model, which was 4.7x the time needed to build Bitcoin’s model. This cost in time commitment is a natural result of working with newly developed and complex systems that provide a rich set of features.

In threat modeling Filecoin’s whitepaper, we found three unaddressed issues, mostly dealing with collusion cases that were not considered. Additionally, there are many places where the system is underspecified and so it is not possible to reason about whether or not it meaningfully addresses a threat.

Ethics and disclosure. We reached out to the Filecoin team, which mentioned efforts they have undertaken to resolve these problems. We withhold details about these issues until later as part of the responsible disclosure process.

5.3 CacheCash Analysis (Steps 1-4)

CacheCash is a cryptocurrency that provides a distributed content delivery service. It allows content providers to construct dynamic networks of caches to serve their clients. This is done by allowing anyone to set up a new cache, and then collect cryptocurrency tokens from publishers for serving their clients. To handle the functionality of the currency exchange medium CacheCash uses a modified version of the Bitcoin protocol. We developed ABC while designing CacheCash, when it became clear that none of the traditional frameworks suited our needs.

As shown in Table 2, the threat model for CacheCash is smaller, in terms of the number of collusion matrices and threat cases, than the one developed for Filecoin. This is because CacheCash provides a single type of service on top of the currency exchange medium, as compared to two in Filecoin.

Beyond threat modeling, we used ABC during the threat mitigation process in the CacheCash design. During that time, we observed the importance of rational financial incentives in the threat mitigation phase. This includes employing detect-and-punish mechanisms in which the penalty deposit of a party is revoked upon detecting that it is cheating, or designing algorithms that, when implemented in a malicious way, can cost the attacker more in resources than would working honestly. Furthermore, we realized the value of game theory and economic analysis in assessing the effectiveness of these economic threat mitigation approaches, and in quantifying the risk, or amount of damage, that financial attacks may cause. To date, we found ABC useful for CacheCash in both the pre-design threat modeling step, and the after-design security analysis of the system modules.

As CacheCash’s design is not public yet, its full threat model will be released at the public unveiling of the system in late 2019.

6 Conclusions

In this paper, we introduce ABC, a cryptocurrency-focused threat modeling framework. Its design is motivated by the observation that traditional threat modeling frameworks do not fit cryptocurrencies, thus leaving them vulnerable to unanticipated attacks. ABC introduces collusion matrices, a technique that allows designers to investigate hundreds of threat cases in a reasonable amount of time. Both the user study and the use cases confirm that our framework is effective in unraveling hidden threat cases. In an on-going work with the Linux foundation, we are applying ABC to cloud native security technologies. This shows the potential of ABC to improve the security of a wide array of distributed systems.


G Deriving ABC Threat Categories

In this section, we show how the threat categories listed in Table 1 are derived. We apply the procedure outlined in Step 2 of the ABC framework (Section 3.2) to the assets of CompuCoin, which include the service, in this case computation outsourcing, service rewards or payments, blockchain, transactions, currency, and communication network. For each asset, we outline its security properties in order to identify any factors that might violate these properties. These factors are labeled as threat categories.

Starting with the service asset, our analysis outlines the following:

  1. Security properties: A secure service can be defined as the action of serving clients correctly at anytime, with properly hiding the service content, and binding the servers to the service they provide. Hence, the security properties of a service asset may include integrity, availability, confidentiality, and non-repudiation.

  2. Threat categories: By negating the above properties, we find that the service asset has the following threat categories:

    • Service tampering/corruption: An attacker provides clients with invalid service or corrupts the correct service delivered by others.

    • Information disclosure: An attacker reveals the contents of service-related messages, such as the service content/outcome, the service requests sent by clients, replies sent by servers, etc.

    • Repudiation: A server denies providing a specific service or a client denies receiving a specific service.

    • DoS: An attacker makes the service unavailable to legitimate users.

Next, we analyze the service payment (or rewards) asset as follows:

  1. Security properties: One may consider the service payment asset secure as long as: a) servers are rewarded properly for the service they provide, and b) servers earn the payments they collect.

  2. Threat categories: Negating the above security requirements produce the following threat categories:

    • Service slacking: A server collects payments without performing all the promised work.

    • Service theft: A client obtains service for a lower payment than the agreed upon amount.

For the blockchain asset, our analysis produces the following:

  1. Security properties: The blockchain security properties are tied to the security of the underlying consensus protocol. These properties have been thoroughly studied in the literature [21, 23, 34]. We adopt the ones introduced in [34] with slight modifications based on the work presented in [21], which include:

    • Consistency: At any point in time, honest miners hold copies of the blockchain that have a common prefix and may differ only in the last blocks, where is a block confirmation parameter. A block then is confirmed once it is buried under blocks on the blockchain.

    • Future-self consistency: At any two points in time, and , the blockchain maintained by an honest party may differ only in the last blocks. Consistency and future-self consistency properties achieve blockchain persistence or immutability.

    • Fairness: Miners collect mining rewards in proportion to the resources they expend in the mining process.

    • Correctness: All the blocks within the longest branch in the blockchain are valid. (Note that correctness and fairness represent the chain quality property outlined in [23, 34].)

    • Growth: As long as the system is functional, new valid blocks will be added to the blockchain.

  2. Threat Categories: By negating the above properties, we can distill the following threat categories for the blockchain asset:

    • Inconsistency: Honest miners do not agree on the prefix of the blockchain copies they hold beyond the unconfirmed blocks. This also covers the case of an honest miner who does not agree with itself on the blockchain prefix it holds over time, e.g., alternating between two branches that compete in being the longest.

    • Invalid block adoption: The longest chain contains corrupted blocks that either have invalid format or contain invalid transactions.

    • Biased mining: A miner pretends to expend the needed resources to be selected to extend the blockchain and collect the mining rewards.

    • Chain freezing: The blockchain does not grow at a regular rate, but instead freezes for several contiguous rounds. This threat category is a form of DoS attack, and hence, we cover it under DoS against the communication network asset.

Next we analyze the transaction asset as follows:

  1. Security properties: Secure transactions can be characterized as correct, tamper-proof, and source-binding, i.e., cannot be denied by the originator. In addition, these transactions need to be accessible to the system users at any time so they can send/receive/view transactions as needed. Moreover, these transactions must not reveal any information about the source, destination, and amount of transferred funds. Accordingly, we outline the following security properties for the transaction asset: non-repudiation, integrity, validity, availability, and anonymity. Note that the validity property is already covered by the correctness aspect of the blockchain, where a valid blockchain contains only valid transactions. Furthermore, the availability property is covered under the communication network asset.

  2. Threat categories: Based on the previous discussion, and again by negating the aforementioned security properties, the threat categories for the transaction asset would be:

    • Repudiation: An attacker denies issuing transactions.

    • Tampering: An attacker manipulates the fields of a transaction.

    • Deanonymization: An attacker violates users’ privacy by exploiting the public nature of the blockchain to link transactions and payments, and use this knowledge to track the activity of these users in the system and, possibly, reveal their real identities.

Next we analyze the currency asset as follows:

  1. Security properties: The security properties of the currency asset are intertwined with the properties of the transaction asset. This is because the currency takes the form of digital tokens, which are the transactions exchanged in the system. Thus, they inherit all the transaction security properties. What remains is to deal with the currency ownership, meaning that only the owner can spend these tokens.

  2. Threat categories: Beside the categories outlined above for the transaction asset, we have the following threat category for the currency asset:

    • Currency theft: An attacker steals currency from others in the system. This includes all currency theft attacks that are not covered by other assets. For example, biased mining is currency theft, where a miner steals others rewards indirectly, but it is already covered by the blockchain. The same holds for the service payment related threats.

Finally, we analyze the communication network asset as follows:

  1. Security properties: The communication network is the backbone of any cryptocurrency system, and one that is unreliable can lead to numerous problems. First, it can create delays in propagating newly mined blocks that could produce an inconsistent blockchain. Second, it can cause delays in relaying transactions, which could reduce the transaction throughput of the system and affect its availability aspect. Third, it can slow down setting up new miners who need a longer time to discover other peers and download copies of the blockchain. Fourth, it opens the possibility of being controlled by external parties that could intercept the communication links and isolate nodes in the network. Consequently, a secure cryptocurrency system needs a reliable and robust communication network. We merge all these aspects into one security property, namely, availability.

  2. Threat categories: The communication network asset has one threat category, which is DoS.

Table 1 summarizes all the threat categories derived in this section. As mentioned previously, this table is by no means comprehensive. Additional threats can be added based on the asset types of the system, or more refined definitions of the asset security properties. This detailed treatment was provided as a thorough example to clarify the application of Step 2 in the ABC framework. Nevertheless, we found this table sufficiently detailed when building threat models for the systems reported as use cases in Section 5, including Bitcoin, Filecoin, and CacheCash, as well as for the user study tutorial as reported in Section 4. For this reason, Table 1 can be viewed as a base threat list that can be extended, or even reduced, based on the system under design.