Hello, my name is Mika Ylianttila, and today I will talk about threat landscape and potential solutions in 6G security, trust and privacy. I'm leading a research group called NETSEC, Network, Security, Trust, and Privacy, and we are focusing on studies developing secure, scalable and resource efficient techniques in 5G and beyond 5G IoT systems. Previously, our work has included a white paper on this area where we have studied the research challenges on this area for the wider community Our research group, as mentioned, focuses on these topics and you can find our web page here.
Okay, so let's go directly to our today's topic. What are the main drivers and emerging technologies in this area? Of course, one of the main things is the emergence of AI technologies, which are based on machine learning widely. Generative AI is widely used and is finding new use cases also in the communications systems. On the other hand, similar machine learning techniques has been used previously already, so it is at the same time kind of old and new thing. But from the security point of view, it means that AI can be now used in two ways. In one way, it can be used as a tool for defence.
In the other way, it can be used for adversarial attacks towards the communication systems. Also, we are witnessing utilization of communication in new verticals and use cases, also in the dual use cases, in the civilian and military use cases, when these new technologies are developed, and they will be widely used also as part of the critical infrastructure of society. One challenge will be the diversity and volume of the new IoT devices connected for the network. And also we are closing in on the time when quantum computing is coming as a more realistic scenario. It will be interesting to see whether it will be reality in the 6G era in the 2030s or not, but it is anyway good to be prepared also for that. Okay, so regarding the security related threats in 6G era.
It will be quite obvious that we will have both inherited and novel threats. The inherited threads comes from the kind of previous generation. It will be a kind of a good educated guess that the first phases of 6G will be using some parts of the 5G network, such as the 5G core network, similar to the evolution we had in the 5G. So 5G started with the non standalone version where we had 4G core network and then new radio in the first phase, and then the standalone version was when we had also the 5G core network. So it could be a good educated guess that 6G could follow similar evolution path.
So in the first phase it would use the 5G core network. So basically all the security vulnerabilities regarding 5G, which are not yet solved before the 6G time will be relevant for 6G security. So of course the major enabling technologies of 5G will be there probably, so this edge computing platform and data integrity, data confidentiality will be important. In addition, the physical and Mac layer of security hardening is very relevant when we move on in the evolution.
And we will need rapid and early threat detection as we already need today. So I think that one of the main thing is that we will need defensible architectures because the current philosophy is that it is very difficult to get the fully secure system, but instead we need to think how we can defend the systems as well as possible. So the next thing is also related here that how we can build trust. The prevailing principle is this zero trust architecture where we don't have really inherited trust zones.
And so these new trust policies and trust enforcement is important in 6G as well. Of course, the privacy is the third facet of this important thing. So especially as we have a very data driven business and also considering AI itself. AI is a very kind of data-driven area. So wherever AI is used, there's a lot of data typically.
So how we can preserve privacy and it will be an important question. And of course, for that regulation is needed. And also new technical solutions such as federated-learning type of solutions, which have been under study in recent years. We have been studying especially hierarchical and fully distributed learning cases and considered model poisoning and inference attacks. Here we can all again see some of the highlights or key points in the security threats landscape in 6G. So one thing to consider today is that today we are living in this pre 6G security and pre-standardization.
We don't yet have the 6G standards. They are coming soon. The standardization work is starting quite soon. But it will be quite sure that these same enabling technologies, as mentioned before, will be under scrutiny.
Software-defined networking, network functions, virtualization, mobile edge computing, and network slicing will be important to consider from the security perspective. Of course, the security of 6G architecture itself will be important: how to build the security there and considering a may be new type of more distributed features in the network architecture. And also more automation will be brought into network using AI as an example. We will need to consider widely the security threats of machine learning, especially, and consider what are the emerging vulnerabilities if we take more automation in the system and then we'll use more widely these certain type of machine learning algorithms. Also, there are scenarios related to quantum-computing-resistant algorithms. So that's something that's also good to consider.
And also some other new enabling technologies like blockchain and visible light communication will need some security consideration. Okay. So what will be needed is this anomaly based intrusion detection, which will be, is it already used in the kind of standard practice, but we need to develop them further to take into account the new features of distributed learning, for example. And our aim is to mitigate security vulnerability impacts with good design and using optimized and robust algorithms to help these matters. At some point, quantum machine learning could be integrated with the learning algorithms.
Okay, so there are a couple of areas which are difficult or challenging. Of course, one big challenge will be that when the AI or machine learning techniques are utilized in an adversarial manner and how we can detect them early and what kind of defences we can build for those. Also other issues, which are important to consider are privacy issues, as mentioned, ethical issues to mention a few.
We have a project where we are studying these topics, mainly focusing on the certain type of attacks. So poisoning attack for model poisoning is the most basic threat and vulnerability in federated learning. So we are studying that how to make robust algorithms for that. Also model inference attack is considered in the security model. We are also thinking of different use cases, which would be relevant for 6G era.
Also in the evolution path to 6G, like 5G advanced systems, like autonomous driving, industrial IoT settings and that kind of use cases which can be simulated in a simulation environment. On the other hand, we are also very much interested in practical security testing in real test environment. We have a excellent test environment here in Oulu: 5G test network, which is evolving towards the 6G test network where we can test in practice certain type of attacks and the mitigation in real test environment. So here we have an example based on our recent journal article, and a previous conference paper, where we are considering defence used in the hierarchical federated learning systems. So, the challenge here is that the cloud server receives only the aggregated model updates, making it difficult to defend.
So here we utilize the clustering and clustering averaging where we can select the best cluster average so that with this mathematical principle, we can use the cluster to proceed with the learning. So, we can protect the system against that the model will be poisoned and the system to learn incorrectly. And we have done some studies with the MNIST dataset, which is a quite well known dataset in the security research. We can protect against 80 percent malicious nodes in the model, which is better than the typical 50 percent level. Also we can identify malicious nodes in the learning system and then remove them from the network to improve the efficiency.
It must be noted that one of the challenges in decentralized learning is really this model poisoning because it is something that we cannot protect only by encryption. So that means that the model can be poisoned before it is sent over the air. So for that reason, we have to have these mechanisms to protect against the model poisoning. So that is very important.
The model integrity cannot be compromised. And we need to consider these different factors overfitting heterogeneous resources. And adjust things that may compromise the integrity of the model.
Also, the kind of architecture there needs to be considered whether we need some level of centralization, how we can take into account that we may have a kind of heterogeneous participants with different roles in the model distribution. As it was mentioned, we are in the zero trust architecture era, so we cannot trust that all members in the system are good willing so this membership inference is an issue we need to consider. Also, already in the training phase of the, kind of the training data inferences need to be considered. So we have a couple of solutions there, which have been considered.
So this kind of gossip learning is one of the relevant techniques utilized already. And also some use cases may come for this kind of distributed ledger technologies to verify some of the transactions. So blockchain type of technologies can be used to verify certain transactions.
So that is important for the zero trust architecture kind of thinking where the principle is that never trust, always verify. And lastly, this kind of privacy-enabling technologies in addition to federated learning include homomorphic encryption, which enables preserving privacy better in certain situation, and that can be considered a case by case. Okay, then a couple of open issues, what we need to also consider. So this, for example, the differential privacy has been expected to emerge and there are some practical implementation existing. But one of the challenges is that it has a considerable privacy performance trade off. So that's something to consider when we already are in the face of 5G facing really tight end to end delay requirements.
And as we are moving forward, the performance considerations are also relevant. So when we are taking into use the more advanced security technologies, we need to also consider these trade offs of performance, energy consumption, and towards developing also sustainable systems. Also regarding these kind of distributed defences, we need to have more white box implementation.
So the open source type of implementations are needed. Examples of these include O-RAN activities where we have transparency of the implementation gives us more ways to evaluate the redistributed and overall transparency of the solutions. Of course, then also important aspects is this, what kind of data sets are used in the research and development. So many times we are using data sets, which are not from the real network data, but they are other sensitized or simulated data.
So creation of new realistic data sets is important and then testing these new algorithms, not only with the synthesized and simulated data, but also in realistic data sets and also realistic environments. So that's when we will finally ultimately see what are the truly working solutions. And also different other concerns are needed to consider regarding the technologies themselves, depending on what kind of technology they are using, neural networks or other similar technologies. Okay, then at the end, a couple of future research directions. Of course, we need to think of the kind of end-to-end design and how we can integrate AI models securely, efficiently, how we can monitor the systems efficiently, and also in the model level, how we can mitigate the potential biases. So as we are talking about the machine-learning model, so there's always a possibility that there are some biases.
Also, there are many kind of emerging new techniques. For example, in the federal learning field. One of the new trend on new areas is this split learning. Where the attack surface is more limited by using partial user data, partial model updates. Also, one important aspect is that how we can handle the conflicting requirements. And for that reason, very important new aspect is that how could we, for example, utilize the explainable AI type of approach to increase the transparency of the solutions and then try to solve the potentially conflicting requirements regarding asynchronous and synchronous updating of the models.
Finally in the conclusion. So overall, we are in a very exciting era and time to look into this new type of algorithms developing generative AI, very big topic and how it will impact the design and use of the 6G era. And before that, what kind of tools we can develop for the security side to defend and also maybe simulate new kind of attacks and then be able to defend them more effectively. So, decentralization is one theme that we see as a trend when we move from 5G to 6G. So, we expect more decentralization in the system.
And that is good because it can bring more fault tolerance, redundancy, and privacy benefits. And federally learned is one great example of that. Regarding this kind of new vertical areas, so we want to utilize these systems in different domains. So that means different levels of security considerations in different domains. Especially if you consider dual use: use case of civilian and military assistance.
So future needs in this area. We want to look for these topics of distributed adaptation of existing poisoning attack defences, early isolation of malicious and colluding participants and preserving privacy while retaining moral integrity and also implementing trust mechanisms, even though there is this zero trust architecture, but how we can verify the trust and build some level of trust based system there. And also balancing these conflicting requirements and solutions and trying to understand what are the new type of vulnerabilities in the system. So we are also looking for hybrid solutions, which are implementation and context dependent. Overall, this is a very exciting and innovative research area, and we are looking for the emergence of this, also the decentralized learning used in the resilient and secure 6G ecosystem. You can find more information from these topics from our recent research papers, which are mentioned here.
I thank you for your time and hope to see you in the future. Thank you.
2025-01-17 09:41