The annual economic impact of cybercrime is estimated to be higher than that of the Drug trade1 and by some estimates it’s twice as much as the economic impact of the 9/11 attack2. While organisations have had focus on IT security for a long time and have spent vast amounts of money, cyber-attacks and news of hacked systems are far from being history. About 6.5M new malware were created across the Internet in the first quarter of 2013 alone3.
Computers and the internet have become indispensable for homes and organisations alike. The dependence on them increases by the day, be it for household users, in mission critical space control, power grid management, medical applications or for corporate finance systems. But also in parallel are the challenges related to the continued and reliable delivery of service which is becoming a bigger concern for organisations. Cyber security is at the forefront of all threats that the organisations face, with a majority rating it higher than the threat of terrorism or a natural disaster.
In spite of all the focus cyber security has had, it has been a challenging journey so far. The global spend on IT security is expected to hit $120 billion by 20174, and that is one area where the IT budget for most companies either stayed flat or slightly increased even in the recent financial crises5. But that has not substantially reduced the number of vulnerabilities in software or attacks by criminal groups.
There is a need to fundamentally rethink our approach to securing our IT systems. Our approach to security is siloed and focuses on point solutions so far for specific threats like anti viruses, spam filters, intrusion detections and firewalls6. But we are at a stage where cyber systems are much more than just tin-and-wire and software. They involve systemic issues with a social, economic and political component.
The interconnectedness of systems, intertwined with a people element makes IT systems un-isolable from the human element. Complex cyber systems today almost have a life of their own; Cyber systems are complex adaptive systems that we have tried to understand and tackle using more traditional theories.
Complex systems - an introduction
Before getting into the motivations of treating a cyber system as a complex system, here is a brief of what a complex system is. Note that the term ‘system’ could be any combination of people, process or technology that fulfils a certain purpose. The wrist watch you are wearing, the sub-oceanic reefs, or the economy of a country - are all examples of a ‘system’.
In very simple terms, a complex system is any system in which the parts of the system and their interactions together represent a specific behaviour, such that an analysis of all its constituent parts cannot explain the behaviour. In such systems the cause and effect can not necessarily be related and the relationships are non-linear - a small change could have a disproportionate impact. In other words, as Aristotle said: ‘the whole is greater than the sum of its parts’.
One of the most popular examples used in this context is of an urban traffic system and emergence of traffic jams; analysis of individual cars and car drivers cannot help explain the patterns and emergence of traffic jams.
A complex adaptive system (CAS) also has characteristics of self-learning, emergence and evolution among the participants of the complex system. The participants or agents in a CAS demonstrate heterogeneous behaviour. Their behaviour and interactions with other agents continue to evolve. The key characteristics for a system to be characterised as complex adaptive are:
- The behaviour or output cannot be predicted simply by analysing the parts and inputs of the system.
- The behaviour of the system is emergent and changes with time. The same input and environmental conditions do not always guarantee the same output.
- The participants or agents of a system (human agents in this case) are self-learning and change their behaviour based on the outcome of the previous experience.
Complex processes are often confused with ‘complicated’ processes. A complex process is something that has an unpredictable output, however simple the steps might seem. A complicated process is something with lots of intricate steps and difficult to achieve pre-conditions, but with a predictable outcome. An often used example is: making tea is complex (at least for me ... I can never get a cup that tastes the same as the previous one), building a car is complicated. David Snowden’s Cynefin framework gives a more formal description of the terms7.
Complexity as a field of study isn’t new, its roots could be traced back to the work on metaphysics by Aristotle8. Complexity theory is largely inspired by biological systems and has been used in social science, epidemiology and natural science study for some time now. It has been used in the study of economic systems and free markets alike and gaining acceptance for financial risk analysis as well. It is not something that has been very popular in the cyber security so far, but there is growing acceptance of complexity thinking in applied sciences and computing
Motivation for using complexity in cyber security
IT systems today are all designed and built by us (as in the human community of IT workers in an organisation plus suppliers) and we collectively have all the knowledge there is to have regarding these systems. Why then do we see new attacks on IT systems every day that we had never expected, attacking vulnerabilities that we never knew existed?
One of the reasons is the fact that any IT system is designed by thousands of individuals across the whole technology stack from the business application down to the underlying network components and hardware it sits on. That introduces a strong human element in the design of cyber systems and opportunities become ubiquitous for the introduction of flaws that could become vulnerabilities9.
Most organisations have multiple layers of defence for their critical systems (layers of firewalls, IDS, hardened O/S, strong authentication etc.), but attacks still happen. More often than not, computer break-ins are a collision of circumstances rather than a standalone vulnerability being exploited for a cyber-attack to succeed. In other words, it’s the ‘whole’ of the circumstances and actions of the attackers that cause the damage.
Reductionism vs Holism approach
Reductionism and Holism are two contradictory philosophical approaches for the analysis and design of any object or system. The reductionists argue that any system can be reduced to its parts and analysed by ‘reducing’ it to the constituent elements; while the holists argue that the whole is greater than the sum so a system cannot be analysed merely by understanding its parts10.
Reductionists argue that all systems and machines can be understood by looking at its constituent parts. Most of the modern sciences and analysis methods are based on the reductionist approach, and to be fair they have served us quite well so far. By understanding what each part does you really can analyse what a wrist watch would do, by designing each part separately you really can make a car behave the way you want to, or by analysing the position of the celestial objects we can accurately predict the next solar eclipse. Reductionism has a strong focus on causality - there is a cause to an affect.
But that is the extent to which the reductionist view point can help explain the behaviour of a system. When it comes to emergent systems like human behaviour, socio-economic systems, biological systems or socio-cyber systems, the reductionist approach has its limitations. Simple examples like the human body, the response of a mob to a political stimulus, the reaction of the financial market to the news of a merger, or even a traffic jam - cannot be predicted even when studied in detail.
We have traditionally looked at Cyber security with a reductionist lens with specific point solutions for individual problems and tried to anticipate the attacks a cyber-criminal might do against known vulnerabilities. It’s time we start looking at cyber security with an alternate holism approach as well.
Computer break-ins are like pathogen infections
Computer break-ins are more like viral or bacterial infections than a home or car break-in9. A burglar breaking into a house can’t really use that as a launch pad to break into the neighbours. Neither can the vulnerability in a one lock system for a car be exploited for a million others across the globe simultaneously.
They are more akin to microbial infections to the human body, they can propagate the infection as humans do. They are likely to impact large portions of the population of a species as long as they are ‘connected’ to each other and in the case of severe infections the systems are generally ‘isolated’; as are people put in ‘quarantine’ to reduce further spread9.
Even the lexicon of cyber systems uses biological metaphors - virus, worms, infections etc. It has many parallels in epidemiology, but the design principles often employed in cyber systems are not aligned to the natural selection principles. Cyber systems rely a lot on uniformity of processes and technology components as against diversity of genes in organisms of a species that make the species more resilient to epidemic attacks11.
Complexity theory has gained great traction and proven quite useful in epidemiology, understanding the patterns of the spread of infections and ways of controlling them. Researchers are now turning towards using their learnings from natural sciences to cyber systems.
Approach to mitigating security threats
Traditionally there have been two different and complimentary approaches to mitigate security threats to cyber systems that are in use today in most practical systems11:
Formal validation and testing
This approach primarily relies on the testing team of any IT system to discover any faults in the system that could expose a vulnerability and can be exploited by attackers. This could be functional testing to validate that the system gives the correct answer as it is expected, penetration testing to validate its resilience to specific attacks, and availability / resilience testing. The scope of this testing is generally the system itself, not the frontline defences that are deployed around it.
This is a useful approach for fairly simple self-contained systems where the possible user journeys are fairly straightforward. For most other interconnected systems, formal validation alone is not sufficient as it’s never possible to ‘test it all’.
Test automation is a popular approach to reduce the human dependency of the validation processes, but as Alan Turing’s ‘Halting problem of Undecideability’1 proves - it’s impossible to build a machine that tests another one in all cases. Testing is only anecdotal evidence that the system works in the scenarios it has been tested for, and automation helps get that anecdotal evidence quicker.
Encapsulation and boundaries of defence
For systems that cannot be fully validated through formal testing processes, we deploy additional layers of defences in the form of firewalls or network segregation or encapsulate them into virtual machines with limited visibility of the rest of the network etc. Other common techniques of additional defence mechanisms are intrusion prevention systems, anti-virus etc.
This approach is ubiquitous in most organisations as a defence from the unknown attacks as it’s virtually impossible to formally ensure that a piece of software is free from any vulnerability and will remain so.
Approaches using complexity sciences could prove quite complementary to the more traditional ways. The versatility of computer systems make them unpredictable, or capable of emergent behaviour that cannot be predicted without ‘running it’11. Also running it in isolation in a test environment is not the same as running a system in the real environment that it is supposed to be in, as it’s the collision of multiple events that causes the apparent emergent behaviour (recalling holism!).
Diversity over uniformity
Robustness to disturbances is a key emergent behaviour in biological systems. Imagine a species with all organisms in it having the exact same genetic structure, same body configuration, similar antibodies and immune system - the outbreak of a viral infection would have wiped out complete community. But that does not happen because we are all formed differently and all of us have different resistance to infections.
Similarly some mission critical cyber systems especially in the aerospace and medical industries implement ‘diversity implementations’ of the same functionality and a centralised ‘voting’ function decides the response to the requester if the results from the diverse implementations do not match.
It’s fairly common to have redundant copies of mission critical systems in organisations, but they are homogenous implementations rather than diverse - making them equally susceptible to all the faults and vulnerabilities as the primary ones.
If the implementation of the redundant system is made different from the primary - a different O/S, different application container or database versions - the two variants would have different levels of resilience to certain attacks. Even a change in the sequence of memory stack access could vary the response to a buffer overflow attack on the variants12 - highlighting the central ‘voting’ system that there is something wrong somewhere.
As long as the input data and the business function of the implementation are the same, any deviations in the response of the implementations is a sign of potential attack. If a true service-based architecture is implemented, every ‘service’ could have multiple (but a small number of) heterogeneous implementations and the overall business function could randomly select which implementation of a service it uses for every new user request. A fairly large number of different execution paths could be achieved using this approach, increasing the resilience of the system13.
Multi variant execution environments (MVEE) have been developed, where applications with slight differences in implementation are executed in lockstep and their response to a request are monitored12. These have proven quite useful in intrusion detection trying to change the behaviour of the code, or even identifying existing flaws where the variants respond differently to a request.
Along similar lines, using the N-version programming concept14, an N-version antivirus was developed at the University of Michigan that had heterogeneous implementations looking at any new files for corresponding virus signatures. The result was a more resilient anti-virus system, less prone to attacks on itself and with 35 per cent better detection coverage across the estate15.
Agent-based modelling (ABM)
One of the key areas of study in complexity science is agent-based modelling, a simulation modelling technique.
Agent-based modelling is used to understand and analyse the behaviour of complex systems, specifically complex adaptive systems. The individuals or groups interacting with each other in the complex system are represented by artificial ‘agents’ and act by predefined set of rules.
The agents could evolve their behaviour and adapt as per the circumstances. Contrary to deductive reasoning* (most popularly used to explain the behaviour of social and economic systems), simulation does not try to generalise the system and agents’ behaviour.
ABMs have been quite popular to study things like crowd management behaviour in case of a fire evacuation, spread of epidemics, to explain market behaviour and, recently, financial risk analysis. It is a bottom-up modelling technique wherein the behaviour of each agent is programmed separately, and can be different from all other agents. The evolutionary and self-learning behaviour of agents could be implemented using various techniques, genetic algorithm implementation being one of the popular ones16.
Cyber systems are interconnections between software modules, wiring of logical circuits, microchips, the Internet and a number of users (system users or end users). These interactions and actors can be implemented in a simulation model in order to do what-if analysis, predict the impact of changing parameters and interactions between the actors of the model.
Simulation models have been used for analysing the performance characteristics based on application characteristics and user behaviour for a long time now - some of the popular capacity and performance management tools use the technique. Similar techniques can be applied to analyse the response of cyber systems to threats, designing a fault-tolerant architecture and analysing the extent of emergent robustness due to diversity of implementation.
One of the key areas of focus in agent-based modelling is the ‘self-learning’ process of agents. In the real world, the behaviour of an attacker would evolve with experience. This aspect of an agent’s behaviour is implemented by a learning process for agents, genetic algorithm’s being one of the most popular techniques for that.
Genetic algorithms have been used for designing automobile and aeronautics engineering, optimising the performance of Formula One cars17 and simulating the investor learning behaviour in simulated stock markets (implemented using agent-based models).
An interesting visualisation of genetic algorithm - or a self-learning process in action - is the demo of a simple 2D car design process (Chrome recommended for viewing) that starts from scratch with a set of simple rules and ends up with a workable car from a blob of different parts.
The self-learning process of agents is based on ‘Mutations’ and ‘Crossovers’ - two basic operators in genetic algorithm implementation. They emulate the DNA crossover and mutations in biological evolution of life forms. Through crossovers and mutations, agents learn from their own experiences and mistakes. These could be used to simulate the learning behaviour of potential attackers, without the need to manually imagine all the use cases and user journeys that an attacker might try to break a cyber system with.
A way to go…
Complexity in cyber systems, especially the use of agent-based modelling to assess the emergent behaviour of systems is a relatively new field of study with very little research done on it yet. There is still some way to go before using agent-based modelling becomes a commercial proposition for organisations. But given the focus on cyber security and inadequacies in our current stance, complexity science is certainly an avenue that practitioners and academia are increasing their focus on.
Commercially available products or services using complexity-based techniques will, however, take a while till they enter the mainstream commercial organisations.
* Deductive reasoning is a ‘top-down’ reasoning approach starting with a hypothesis and data points used to substantiate the claim. Inductive reasoning on the other hand is a ‘bottom-up’ approach that starts with specific observations which are then generalised to form a general theory
- 1 A. Lewis and S. Baker, ‘The Economic Impact of Cybercrime and Cyber Espionage,’ 22 July 2013. (Online). (Accessed 17 June 2014).
- 2 Kugel, ‘Terrorism and the Global Economy,’ E-International Relations Students, 31 Aug 2011. (Online). (Accessed 17 June 2014).
- 3 ‘Cybersecurity - Facts and Figures,’ International Telecommunications Union, (Online). (Accessed 11 June 2014).
- 4 ‘Interesting Facts on Cybersecurity,’ Florida Tech University Online, (Online). (Accessed 12 June 2014).
- 5 ‘Global security spending to hit $86B in 2016,’ 14 Sep 2012. (Online).
- 6 Forrest, S. Hofmeyr and B. Edwards, ‘The Complex Science of Cyber Defence,’ 24 June 2013. (Online).
- 7 ‘Cynefin Framework’ (David Snowden), (Online).
- 8 ‘Metaphysics’ (Aristotle), (Online).
- 9 Armstrong, ‘Motivation for the Study and Simulation of Cybersecurity as a Complex System,’ 2008.
- 10 A. McLeod, ‘Reductionism and Holism’, 2008.
- 11 C. Armstrong, J. R. Mayo and F. Siebenlist, ‘Complexity Science Challenges in Cybersecurity,’ March 2009.
- 12 Salamat, T. Jackson, A. Gal and M. Franz, ‘Orchestra: Intrusion Detection Using Parallel Execution and Monitoring of Program Variants in User-Space,’ Proceedings of the 4th ACM European conference on Computer systems, pp. 33-46, April 2009.
- 13 C. Armstrong and J. R. Mayo, ‘Leveraging Complexity in Software for Cybersecurity (Abstract),’ Association of Computing Machinery, pp. 978-1-60558-518-5, 2009.
- 14 Liming and A. Avizienis, ‘N-VERSION PROGRAMMINC: A FAULT-TOLERANCE APPROACH TO RELIABILITY OF SOFTWARE OPERATION,’ Fault-Tolerant Computing, p. 113, Jun1995.
- 15 Oberheide, E. Cooke and F. Jahanian, ‘CloudAV: N-Version Antivirus in the Network Cloud,’ University of Michigan, Ann Arbor, MI 48109, 2008.
- 16 H. Holland, ‘Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence’, Michigan: University of Michigan Press, 1975.
- 17 B. P. J. Wloch, ‘Optimising the performance of a formula one car using a genetic algorithm,’ Parallel Problem Solving from Nature-PPSN VIII, pp. 702-711, January 2004.
- 18 (. o. D. Leon, ‘Press Transcript,’ US Department of Defence, 11 Oct 2012. (Online). (Accessed 12 June 2014).