Cyber security solutions for smart traffic control systems
Lead Research Organisation:
University of Southampton
Department Name: Sch of Electronics and Computer Sci
Abstract
We aim to develop a solution framework that can efficiently tackle the cyber security vulnerabilities of smart traffic control systems. Such solutions would be very beneficial for both the industry and society, as current Internet of Things (IoT) based networks, which enjoy significant interest as a key technology towards the development of smart societies, are typically very vulnerable against cyber attacks. This is especially true in the case of traffic systems, as they are among the key national strategic infrastructures that have to be protected at the highest security level.
This proposal is the first study that aims to propose a solution to these cyber security challenges. In particular, we believe that such efficient solutions need to meet the following criteria:
1. Human-agent heterogeneity: The system design has to be capable of dealing with the heterogeneity of interactions and information sharing between different participating devices (i.e., agents) and humans within the traffic control system.
2. Robustness: the system has to be resistant against different major attack scenarios that may significantly vary in both motives and execution. Adversaries could be very strategic in planning sequential attack actions.
3. Adaptivity: The defence mechanisms should react in real time, in an online manner, and adaptive to the concrete actions of the attackers.
4. Limited resources: The defence strategies have to take into account that there are typically limited resources available to execute each step, or decision made during the process.
To address these criteria, we propose a solution that relies on three fundamental components: (i) human-agent collectives based framework; (ii) game theory; and (iii) resource-constrained online machine learning. In particular, the human-agent collectives based perspective allows us to build a unified framework for smart traffic control systems that can efficiently deal with the heterogeneity between different participating agents and humans. We then use security game theory to discover and analyse major attack scenarios, in order to make our system design robust against them. We then apply resource-constrained online machine learning algorithms to develop efficient adaptive and real-time defence mechanisms. Additionally, these mechanisms also consist of further game theoretic approaches that can be used to predict the strategic behaviour of the attackers, and thus, can make more efficient decisions. Finally, we will build a real testbed as a proof-of-concept of our proposal.
There are many advantages of this solution, namely:
1. The unified human-agent collectives based framework will enable us to define formal descriptions and categorisations of interactions within the system in a principled manner. This will provide a strong basis to develop methodologies to identify and analyse suspicious behaviours.
2. With the game theoretic framework, we will be able to investigate the worst case scenarios of different attack types. This will provide us a full understanding of what should and should not be taken into account during the design of the system.
3. The game theoretic framework also provides an efficient tool to analyse the strategic behaviour of the attackers. This knowledge will play an essential role in predicting the possible future actions of the attackers.
4. The online defence mechanism combines the extracted knowledge, provided by the strategic behaviour analysis within the human-agent collectives and game theoretic frameworks, with efficient machine learning techniques to quickly and efficiently detect malicious behaviours.
5. Given the available resources, the defence mechanism then can decide what is the best response actions that can either prevent the malicious user to cause any harm (in case of sufficient resources are available), or to minimise the damage the attacker can cause (in case of having restricted resources).
This proposal is the first study that aims to propose a solution to these cyber security challenges. In particular, we believe that such efficient solutions need to meet the following criteria:
1. Human-agent heterogeneity: The system design has to be capable of dealing with the heterogeneity of interactions and information sharing between different participating devices (i.e., agents) and humans within the traffic control system.
2. Robustness: the system has to be resistant against different major attack scenarios that may significantly vary in both motives and execution. Adversaries could be very strategic in planning sequential attack actions.
3. Adaptivity: The defence mechanisms should react in real time, in an online manner, and adaptive to the concrete actions of the attackers.
4. Limited resources: The defence strategies have to take into account that there are typically limited resources available to execute each step, or decision made during the process.
To address these criteria, we propose a solution that relies on three fundamental components: (i) human-agent collectives based framework; (ii) game theory; and (iii) resource-constrained online machine learning. In particular, the human-agent collectives based perspective allows us to build a unified framework for smart traffic control systems that can efficiently deal with the heterogeneity between different participating agents and humans. We then use security game theory to discover and analyse major attack scenarios, in order to make our system design robust against them. We then apply resource-constrained online machine learning algorithms to develop efficient adaptive and real-time defence mechanisms. Additionally, these mechanisms also consist of further game theoretic approaches that can be used to predict the strategic behaviour of the attackers, and thus, can make more efficient decisions. Finally, we will build a real testbed as a proof-of-concept of our proposal.
There are many advantages of this solution, namely:
1. The unified human-agent collectives based framework will enable us to define formal descriptions and categorisations of interactions within the system in a principled manner. This will provide a strong basis to develop methodologies to identify and analyse suspicious behaviours.
2. With the game theoretic framework, we will be able to investigate the worst case scenarios of different attack types. This will provide us a full understanding of what should and should not be taken into account during the design of the system.
3. The game theoretic framework also provides an efficient tool to analyse the strategic behaviour of the attackers. This knowledge will play an essential role in predicting the possible future actions of the attackers.
4. The online defence mechanism combines the extracted knowledge, provided by the strategic behaviour analysis within the human-agent collectives and game theoretic frameworks, with efficient machine learning techniques to quickly and efficiently detect malicious behaviours.
5. Given the available resources, the defence mechanism then can decide what is the best response actions that can either prevent the malicious user to cause any harm (in case of sufficient resources are available), or to minimise the damage the attacker can cause (in case of having restricted resources).
Planned Impact
Knowledge: As our proposed research is the first to tackle the cyber security challenges of smart traffic control systems, it will enable significant benefits to the traffic control systems research and development community, as well as to wider communities such as Internet of Things (IoT), cyber security, and smart societies. In particular, our findings will provide a generic framework that will allow us to design cyber security solutions for IoT systems in a principled manner, making the dream of building smart societies in a safer way.
Economy and Society: The proposed research will be of significant benefit to our industrial partners. In particular, Sairui-Tech has explicitly identified its benefits from our findings in the corresponding Letter of Support. In addition, while other partners such as MRG Effitas and Intellisense.IO have not provided explicit evidences of direct benefits, they have expressed their interest to integrate our findings into their products. On the other hand, the society will also benefit from our findings, as in the presence of secure solutions, people can trust IoT based system better, and thus, increasing the speed of transition to smart societies.
People: Throughout the project, the research assistants (RA) and the software engineering technician will gain significant expertise in designing efficient cyber security solutions for IoT systems. This expertise is highly sought after in both industry and academia, since tendencies show that IoT systems will be one of the main technologies of the future, which will need high quality cyber security solutions.
Economy and Society: The proposed research will be of significant benefit to our industrial partners. In particular, Sairui-Tech has explicitly identified its benefits from our findings in the corresponding Letter of Support. In addition, while other partners such as MRG Effitas and Intellisense.IO have not provided explicit evidences of direct benefits, they have expressed their interest to integrate our findings into their products. On the other hand, the society will also benefit from our findings, as in the presence of secure solutions, people can trust IoT based system better, and thus, increasing the speed of transition to smart societies.
People: Throughout the project, the research assistants (RA) and the software engineering technician will gain significant expertise in designing efficient cyber security solutions for IoT systems. This expertise is highly sought after in both industry and academia, since tendencies show that IoT systems will be one of the main technologies of the future, which will need high quality cyber security solutions.
Publications
Qingyu Guo
(2017)
Playing repeated network interdiction games with semi-bandit feedback
Youzhi Zhang
(2017)
Optimal escape Interdiction on transportation networks
Marcin Waniek
(2017)
The dollar auction with spiteful players
| Description | The main goal of the project is to provide efficient defence mechanisms against a set of major attacks against both current and future smart traffic control systems. In particular, we look at 3 different levels of attacks: 1. device level; 2. local control module level; and 3. system level. Here are our recent research findings regarding each level of attacks: 1.1. At the device level, we looked at data poisoning type attacks. In particular, we have successfully revealed major vulnerability points of outlier detection algorithms, that are typically used as first line defence mechanisms in traffic control systems (e.g., to detect strange behaviours in the traffic). This result was presented at GameSec 2016 conference as a poster, a leading conference on applying game theory to security. We also aim to submit a more detailed version to a top smart transportation conference (e.g., IEEE ITSC) in 2017. 1.2. We have also proposed an efficient defence mechanism against the abovementioned attack against outlier detection methods. This mechanism is a non-trivial combination of game theoretical and machine learning models. We have put these findings into a technical report, which will be submitted to a top AI/Security conference in the near future. 2.1. We have also investigated a local control level attack, in which the attacker can maliciously modify the traffic light of certain parts of the system to manipulate the traffic. Our findings have been submitted to a major AI conference. 2.2. We also looked at attacks against smart traffic algorithms that use clustering as an AI technique to efficiently set traffic lights. This is a work in progress and we aim to have a publishable version by the end of the summer. In particular, an undergraduate student is working on this topic as his year-3 project. The current status is that we already have the clustering and the attack mechanisms implemented, with a user friendly GUI to simulate the attacks. For the next step, we aim to develop efficient defence mechanisms against these attacks. 3.1. For system level attack scenarios, we looked at a resource allocation dilemma, called the Dollar Auction, in which both the defender and attacker has limited resources, and can decide how much they should use for the defence/attack. Our findings were published at AAAI 2017, the top artificial intelligence conference (acceptance rate is less than 25%, with more than 2500 submissions). |
| Exploitation Route | Our findings could form a basis for further research in the area. In particular, our work provides a principled way of analysis for security problems in smart traffic control. Hence, we believe our results can lead to both theoretically sound and practically efficient solutions in the field. |
| Sectors | Digital/Communication/Information Technologies (including Software),Security and Diplomacy,Transport |
| Description | Academic impact: The project has formed the baseline for 2 PhD projects on machine learning with strategic players. Both are funded by EPSRC (Dinh, Bishop). The findings of the project have also initiated a new line of research, namely: last round convergence with manipulating the game's payoff matrix. This has been resulted in a number of follow-up publications (Bishop, Tran-Thanh, and Gerding, NeurIPS 2021; Dinh et al., ALT 2021; Bishop, Dinh, and Tran-Thanh, AAMAS 2021; Rangi et al., AAAI 2022) and a number of other research papers under submission. Industrial impact: -Dinh has spent 6 months at Huawei as a PhD intern in 2021, and has initiated a new research collaboration with Huawei on last round convergence (as a follow-up topic from the project). -Rangi has moved to Amazon, also in 2021, and since then has also started a new research project on manipulating online learning models. -Tran-Thanh has started a new collaboration with FSOFT QAI, an international software development company (headquartered in Vietnam), to set up a new research group working on manipulating reinforcement learning models. |
| First Year Of Impact | 2021 |
| Sector | Digital/Communication/Information Technologies (including Software) |
| Impact Types | Economic |
| Description | Bo An Qingyu repeated interdiction games |
| Organisation | Nanyang Technological University |
| Department | School of Chemical and Biomedical Engineering |
| Country | Singapore |
| Sector | Academic/University |
| PI Contribution | We started working on the repeated interdiction games as part of the current project. Our team worked on the design of the online learning algorithm. |
| Collaborator Contribution | Bo An's team was working on using the online learning algorithm to solve the problem. |
| Impact | Scientific paper submission to IJCAI 2017, the top AI conference. |
| Start Year | 2016 |
| Description | Waniek Michalak on Dollar Auctions |
| Organisation | University of Warsaw |
| Department | Faculty of Mathematics, Informatics and Mechanics (MIM) |
| Country | Poland |
| Sector | Academic/University |
| PI Contribution | We worked together on a AAAI 2017 paper, with half of the contributions of the theoretical analysis are from our team. In particular, we formulated the problem and provided the analysis for auctions with weakly spiteful players. |
| Collaborator Contribution | Waniek and Michalak worked on the strongly spiteful cases. |
| Impact | Conference publications at AAMAS 2016 and AAAI 2017. |
| Start Year | 2016 |