Digital twins as a security risk?

Abstract

Digital twins have become a popular and widely used tool for assessing risk and resilience, particularly as they have increased in the fidelity and accuracy of their representation of real‐world systems. Although digital twins provide the ability to experiment on and assess risks to and from a system without damaging the real‐world system, they pose potentially significant security risks. For example, if a digital twin of a power system has sufficient accuracy to allow loss of electrical power service due to a natural hazard to be estimated at the address level with a high degree of accuracy, what prevents someone wishing to lead to disruption at this same building from using the model to solve the inverse problem to determine which parts of the power system should be attacked to maximize the likelihood of loss of service to the target facility? This perspective article discusses the benefits and risks of digital twins and argues that more attention needs to be paid to the risks posed by digital twins.

1. INTRODUCTION

A digital twin is a computer‐based representation of a physical system that is typically used for research or planning purposes to assess, for example, risks to the physical system or the impacts of different ways of operating the physical system (e.g., Zio & Miqueles, ²⁰²⁴). A digital twin is thus a model of a system but typically has the connotation of typically being a high‐fidelity, accurate model. Digital twins are attractive because they provide a basis for better understanding the operation of a system, how that system might fail, and how different types of external events may impact that system without experimenting on the actual system and without having data from the real system's performance under the specific scenario(s) of interest. Because of this and the increasing complexity of technical systems, digital twins are an increasingly important part of risk analysis.

Although the term digital twin is relatively new, the basic idea has been around for several decades, at least in engineering. For example, the US nuclear stockpile has been managed based, at least in part, on computer models of nuclear detonations since the last nuclear test in 1992. These models could now be termed digital twins. Similarly, for infrastructure systems, physics‐based computer models of the performance of systems have existed for decades. Examples include hydraulic models of water systems and power flow models for power transmission and distribution systems.

Two key changes have happened in recent years that make digital twins more prevalent and useful while also potentially increasing the security risk they pose. The first is that they have increased in both fidelity and accuracy. The second is that methods have been developed to create highly accurate digital twins of some types of systems, notably infrastructure systems, based on only publicly available data. As will be discussed below, these advances increase the potential for digital twins to be used for nefarious purposes.

The goal of this perspective article is to discuss the potential security risks posed by digital twins and what might feasibly be done to manage this risk. The article starts by briefly summarizing what a digital twin is, how they are used, and what they are used for. It then discusses the potential for reverse engineering a digital twin, that is, using it to solve the inverse problem of determining how best to attack or otherwise hamper the performance of the system. Finally, the article closes with a discussion of what can be done to manage this security risk and a call for more research in this area.

2. DIGITAL TWINS

The term “digital twin” has many definitions. At its most basic level, a digital twin is a computer‐based representation of a physical system that is used for research, planning, or management (often in real‐time) purposes (Zio & Miqueles, 2024). The term probabilistic digital twin is sometimes used when the model in question is a probabilistic model (e.g., Agrell et al., ²⁰²³; Ayello et al., ²⁰²¹). Some digital twins are used in real‐time to support of operation of a complex technical system. Other digital twins are used offline to support, for example, long‐term planning and risk analysis for the system. To better describe digital twins and to better illustrate potential security risks, we use as a specific example of a digital twin the synthetic power system model from Zhai et al. (2021).

Zhai et al. (2021) developed a method for creating a synthetic version of a power distribution system based only on publicly available data. The explicit goal of this approach is to probabilistically estimate which individual buildings in a community will lose electrical power due to a given hazard event. The approach develops a synthetic layout of a power distribution system for a community, including, but not limited to, substations, utility poles, and power lines. It includes an estimation of which power lines are overhead and underground. Others have also developed digital twins of power systems, though they have focused on system operation rather than estimating failures (e.g., Birchfield et al., ²⁰¹⁶; Soltan et al., ²⁰¹⁸). Digital twins such as this have very high degrees of spatial detail, and their accuracy in estimating power outages (Zhai et al., 2021) and power outage durations (Zhai et al., 2022) due to natural hazards has proven strong in holdout validation (Zhai et al., 2021, 2022).

Many other types of digital twins exist, including digital twins of autonomous vehicles (Almeaibed et al., 2021), sewer systems (Bartos & Kerkez, 2021), buildings (Hosamo et al., 2022), and hospitals (Peng et al., 2020), among many others.

3. POTENTIAL FOR INVERSE MODELING AND ASSOCIATED SECURITY RISKS

Inverse modeling is, in a broad sense, using a model initially designed for estimating output given a set of inputs to instead infer what inputs to the model would lead to specific output(s) from the model. An example in the context of our illustrative digital twin could be using the power system digital twin to estimate which combinations of component failures would lead to a loss of power at a given facility.

There are very valid uses of type of inverse modeling for risk assessment. For example, consider a facility serving a critical societal function (e.g., a hospital or fire station). Inverse modeling based on a digital twin could be used to determine which sets of components are most likely to cause a power outage at this facility due to failure during a disruptive event. This can then help facility managers, power system operators, and civil defense planners decide which parts of the power system to reinforce or provide redundancy for, a valuable contribution to risk management for this critical facility.

Now consider instead a different potential use of inverse modeling, this time by an individual or group that has the goal of disrupting the service provided by that facility. Perhaps they are planning other disruptions in the area and want to limit the ability of the community to respond, or perhaps the facility itself is providing such a critical service that disrupting it is their goal. How might this individual or group leverage a digital twin if they had access to it?

If an intelligent adversary had access to a detailed and accurate digital twin, they potentially could use it to determine the best places at which to disrupt the system. To be specific, consider our power system digital twin example. If the attacker did not have the detailed performance model of the system, they may not be able to determine the best places to act to cut power to the facility and may think they need to try to disable more feeders than they would actually need to, leading to a much more difficult attack. However, with a digital twin, they potentially could determine that only one much simpler attack is needed. This is a relatively simple example of using a digital twin to determine a better way to attack a system. A more complicated example would be using a digital twin of a drinking water system to determine where and how to introduce a contaminant to lead to maximum exposure in a population, as in Torres et al. (2009).

As shown in the brief discussion above, digital twins potentially pose a security risk for the systems they are twins of. But what are the necessary and sufficient conditions for a digital twin to pose a security risk? We posit that the following three conditions are necessary and sufficient for a digital twin to pose a security risk. These are based on our experience with digital twins and intelligent adversary risk analysis, but we do not claim that this is an exhaustive list.

• 1) The digital twin must provide sufficiently accurate and detailed estimates of system performance as a function of the inputs to the system.
•  If the digital twin does not accurately estimate system performance as a function of the inputs or the states of the constituent parts of the system, then inverse modeling is unlikely to provide accurate estimates of how the system could be attacked. Similarly, the estimates must be detailed enough in space and/or time, depending on the system and problem, to allow the adversary to estimate precisely where and when to disrupt the system to obtain their desired overall effect. There is admittedly a challenge here. Even a mildly accurate digital twin could lead to an attack against the system that, while not maximizing damage, still does damage. The estimates of system performance thus do not need to be extremely accurate, but they must provide better estimates of the impacts of an attack than the attacker would have been able to create without the digital twin if the digital twin is to have value to the attacker.
• 2) There must be a justified belief that one or more individuals or groups could be interested in attacking or otherwise hampering the performance of the real‐world system and have the capabilities to do so.
•  If the real‐world system for which the digital twin was created is not of interest to potential attackers or if they do not have the capability to carry out the attack suggested by the digital twin, there is likely minimal security risk. For example, one could create a digital twin of a personal bicycle used for commuting. This could then be used to assess how changes in the bicycle or how it is used might impact the commute. Unless someone was interested in disrupting the commute of the bicycle user, there would be minimal security risk associated with this digital twin. Similarly, a digital twin of a power system might lead to a suggested attack against a very well‐protected, inaccessible component of the system. If the attacker does not have the capabilities to overcome these defenses, then there would be minimal security risk associated with the digital twin.
• 3) The digital twin model must be accessible to and usable by the adversary.
•  Implicit in this is that the adversary must know that the digital twin exists and is accurate enough for their purposes. If the adversary does not know the digital twin exists, does not have access to the digital twin, or does not know how to operate the digital twin to achieve the goal of determining how to disrupt the system, the security risk associated with the existence of the digital twin is greatly reduced.

If all three of these conditions are met, there is a security risk posed by the existence of a digital twin. These conditions can be derived for the digital twin setting from the capacity (conditions 1 and 3) and intention (condition 2) aspects commonly referred to when assessing security risk.

One key characteristic of digital twin research is that many digital twins are developed by academic researchers to be accurate and detailed in their estimation. The nature of academic research is that we, as researchers, strive to improve the accuracy and fidelity of digital twin models, both to advance digital twin methodology and to better support decision‐makers using the resulting digital twins.

A second key characteristic of digital twin research is that open publication and, increasingly, open‐sourcing code are highly valued in academic research. There are valid reasons for this. Open access to methods, code, and data is a key part of advancing scientific research. This allows other researchers to more easily build on the methods. It allows potential users of the models to access them without expensive commercial licenses. It increases the transparency of research. Increasingly, open access to data and code is also required by funding agencies.

These two characteristics of digital twin research pose significant challenges from a security perspective. The increasing fidelity and accuracy and increasing move toward open access to digital twin models and data mean that increasingly, two of the necessary conditions for digital twins to pose a security threat are met.

This is not a hypothetical problem. One of the authors of this article was involved in the creation of the Zhai et al. (2021) digital twin model. He was contacted by someone from another country asking for access to the model. That other country is now actively targeting the power system of a neighboring country to try to disable it. Was this a legitimate request or was this an attempt at gaining access to the model to try to solve the inverse problem? It is not clear, but it raises questions. Where there are security concerns, the researchers involved should report these to the appropriate authorities.

4. IMPLICATIONS AND RESEARCH NEEDS

We cannot and will not stop using digital twins. Their benefits for understanding complex technical systems and supporting risk analysis for these systems are too great. However, we as a research community need to take digital twin security risk seriously. How can we continue the development of digital twin methodology while managing the security risk? How should the potential security risk impact how we, as researchers, disseminate our results? This section discusses each of these critical questions in turn.

How can we continue the development of digital twin methodology while managing the security risk? Digital twin research is important and should continue. However, researchers must be cognizant of the security implications of digital twin research. This comes at all phases of the research. When selecting systems to develop digital twins for, researchers should conduct at least a basic assessment of potential security risks if they produce a digital twin of that system. Depending on the outcome of this risk assessment, measures should be implemented to ensure the security of the digital twin model during its development, including both the security of the computational systems used and the security of who has access to the model. Clear data use agreements and data security policies are needed when potentially sensitive data are used to create digital twin models. For digital twins with high‐security risk, a review of the security policies from outside the research team would likely be warranted. These types of procedures are required for some types of systems and data in the United States (e.g., systems and data designated as Critical Energy Infrastructure Information by the Department of Homeland Security), but not for others (e.g., most drinking water systems). Along with this, research is needed on how journals publishing digital twin research that involves potential security risk could insist on transparency and say that security measures are in place without providing so much detail on these that the security is compromised. Digital twin research can be viewed as a special case of research on dual‐use (civilian‐military use) technology. However, what sets digital twin research apart is that it is generally being conducted for types of systems that are not generally thought of as being dual‐use, such as sewer, drinking water, and power systems. Still, there is much that can be learned from how dual‐use technology development is managed.

How should the potential security risk impact how we as researchers disseminate our results? Digital twin researchers have a responsibility to consider security in the dissemination of their results, particularly if the three necessary conditions from Section 3 are met. If these conditions are met, results should, in most cases, not be shown in a journal article or conference presentation in a level of spatial or temporal scale that would allow a potential adversary to use those results to harm the system. This constrains our ability to demonstrate the increasing spatial and temporal accuracy of digital twins. Instead, we are limited to aggregate performance measures as those shown in Zhai et al. (2021) for security reasons. This does potentially still reveal to potential adversaries that higher detail estimates exist, at least if this is discussed in the article. This raises the possibility of the adversary pursuing other means to obtain these, and appropriate security precautions are needed for the model files.

A more challenging question is the degree to which digital twin models should be made openly available as open‐source data and code. The authors’ perspective is that open access to data and models is, in general, valuable to science. It aids others in building from past research and provides full transparency and the ability to confirm past work. However, there are cases in digital twin research where this would not be a prudent course of action. Providing open access to the model code would make it much easier for an adversary to gain access to and use the model to determine how best to disrupt the target system. Even the approach of making the model publicly available but not the input parameter values poses a potential security risk. The attacker could attempt to use publicly available information about the system performance in the past to determine parameter values that approximate the system performance well enough to plan an attack on that system.

There are cases where clearly a digital twin needs to be kept highly secure (e.g., a digital twin of nuclear warhead detonation) and cases where clearly a digital twin could be made open source (e.g., a digital twin of a bicycle used for commuting). In between is a large gray area. Who should be the arbiter of whether or not a digital twin for a given system can be made open source? Currently, this is ad hoc. In many cases, the individual researcher makes this decision with little external guidance. In some cases, such as military‐sponsored research, the sponsor may impose restrictions, but this is currently the exception. Is a more organized process needed? What this would look like and who would be responsible for it, if it is warranted at all, are open questions. In the meantime, one concrete step that journals and conferences can take is to allow exceptions to requirements for open access to digital twin data and code when there is a valid security concern.

One final point to be made is the possibility of using digital twins for digital red teaming. In a traditional red teaming exercise, the red team attacks the system in an effort to find weaknesses, and this is then used to strengthen the defense of the system. This concept could be extended to include the use of digital twins in red teaming. For example, a traditional red teaming exercise could be conducted without the red team having the digital twin. Then the exercise could be repeated with the red team having access to the digital twin. This would allow the system operator to better understand the potential benefit of the digital twin to potential attackers, thus helping them better understand the needed level of security for the digital twin. It would also help the system operator better understand how an adversary might optimally attack their system. This would, in turn, strengthen the system defense.

5. SUMMARY

Digital twins are an invaluable part of risk analysis for complex systems. They allow systems to be understood and experimented with in ways that are not possible based on only the actual physical system and historic data from the physical systems. However, digital twins can, in some situations, pose a security risk. The potential to use inverse modeling based on a digital twin to determine how best to disrupt a system is a very real concern. Those developing and using digital twins have a responsibility to consider this security risk as they plan and develop digital twins, publish and disseminate the results of digital twin research and development, and transition these systems into practical use. We can and should continue to conduct research on digital twin methodology and continue to use these models in practice. However, we must be increasingly thoughtful in how we consider the potential security consequences of this work.

Guikema, S. , & Flage, R. (2025). Digital twins as a security risk? Risk Analysis, 45, 269–273. 10.1111/risa.15749