As compute and intelligence move to the edge to process and analyze data closer to where it is generated, ensuring the trustworthiness of data, compute, and connectivity in a disaggregated, distributed, and multi-stakeholder edge environment becomes paramount. Intel Labs conducts research to address a diverse range of edge and network security challenges, including confidentiality, integrity, accountability, availability, and data privacy. At the same time, it is being collected, stored, transported, and processed. Example areas of research focus include:
- Foundational security technologies such as hardware-based Trusted Execution Environments (TEES) to ensure confidentiality and integrity of data in use.
- Coordinating and establishing trust in a disaggregated and multi-stakeholder edge environment.
- Protocol enhancements and software solutions to mitigate security vulnerabilities in networks.
- Approaches to ensure the trustworthiness of Artificial Intelligence (AI) solutions and workloads targeting the networking domain. Additionally, using AI, such as Large Language Models, to enhance the security of networks.
- Post-quantum encryption techniques to ensure quantum-safe transport of data.
Confidential Computing for Networks
Confidential computing protects data-in-use by performing computations in hardware-based Trusted Execution Environments. Only authorized software or parties can access the data. This secure environment helps businesses realize more value from private, sensitive, or regulated data while remaining protected and compliant. Intel offers two complimentary TEEs, Intel® Software Guard Extensions (Intel® SGX) for process-based enclaves and Intel® Trust Domain Extensions (Intel® TDX) for VM-based enclaves. TEEs and attestation mechanisms provided by them are pivotal for enhancing the security of Virtual Network Functions (VNFs), including the User Plane Function (UPF) in 5G infrastructures. These technologies not only offer a protected execution environment for VNFs but also ensure the legitimacy of the running software through attestation to establish trust over software systems. Notably, our approach achieved a performance close to 93-95% of standard Aether UPF deployment, showcasing that heightened security doesn't necessarily translate to a substantial compromise on performance. Intel Labs' research on end-to-end 5G security emphasizes the integration of these components, establishing a holistic security model that spans from edge devices to core network elements, highlighting Intel's dedication to fostering a comprehensive security framework for future networks.
Multiparty Trust Coordination
Multiparty Trust Coordination (MPTC) enables establishing trust in complex computing environments, such as cloud and edge. For example, it helps relying parties or users increase confidence that a service meets the required security properties when managing sensitive data. However, environment heterogeneity and context-dependent trust requirements cause challenges. For instance, the information describing hardware and software features (for example, platforms, accelerators, code libraries, and databases) can come from multiple sources using different formats and semantics, while trust requirements may involve temporal, geo-location, personal, and institutional aspects. MPTC streamlines the generation, organization, and analysis of such information. Also, by relieving applications from this task, MPTC simplifies their development and maintenance.
Trustworthy AI in Networks
Artificial Intelligence/Machine Learning (AI/ML) solutions are widely expected to be integral to designing, deploying, and operating 5G and 6G networks. The 3rd Generation Partnership Project (3GPP) and Open Radio Access Network (O-RAN) Alliance standards groups are developing several AI-based solutions for intelligent control and management of the core and radio network. For example, several xApps and rApps are being developed for the RAN Intelligent Controller (RIC) in ORAN, which cover applications such as positioning and environment mapping, intelligent radio resource management, and anomaly detection using traffic features. The 3GPP also supports a Network Data and Analytics Function (NWDAF) for network analytics.
Ensuring the robustness of AI solutions to anomalous or malicious inputs is paramount to supporting mission-critical applications in next-generation networks. Intel Labs has extensively worked on showcasing vulnerabilities of AI models and developing suitable defense mechanisms as part of the DARPA GARD (Guaranteeing Adversarial Robustness Against Deception) program, which primarily focuses on vision applications. Intel Labs has also released the open source Modular Adversarial Robustness Toolkit (MART), a unified framework that enables researchers to easily compose novel attacks and defenses for “adversarially robust” deep learning models.
Our recent work has highlighted similar vulnerabilities in AI models currently being developed for radio network control, and we are developing defense mechanisms to improve and certify their robustness to ensure trustworthy AI in wireless networks.
Secure Time-Sensitive Networking
Time-Sensitive Networking (TSN) is critical for coordinating actions in distributed systems. In the case of industrial systems, synchronizing sensing, control, and actuation along a common time axis is crucial for real-time operation. TSN is a set of standards providing determinism for networks through time synchronization based on IEEE 1588 Precision Time Protocol (PTP, IEEE 802.1AS), time-aware traffic scheduling (IEEE 802.1Qbv), and others. Time synchronization entails a clock leader distributing time to the clock followers in the network who update their local system time.
TSN can present a new attack surface where time synchronization can be compromised, resulting in system downtime with financial and safety implications. Conventional security protocols such as Media Access Control Security (MACSec) address some of the threats. However, they do not address attacks that, for example, introduce a delay to disrupt time synchronization. Intel Labs is pursuing a three-pronged research approach comprised of attack detection, localization to identify the source of attack, and recovery from attacks. In the area of attack detection, we cover both attacks that target the clock follower’s platform and attacks on the network switches that cause a drift in the clock follower’s time. The overall goal is to increase the robustness of TSN to adversarial attacks so that it can be used reliably for critical distributed applications.
Enhancing Security Protocols
As networks become more complex, new security threats are revealed while many traditional attack surfaces remain. Our Intel Labs teams are working together to enhance security protocols across the following areas:
Security in a disaggregated world: The opening of closed interfaces plus, the availability of open-source implementation and disaggregation of the RAN and core networks provides more flexibility and easy means of scaling and reconfiguring the networks. However, these improvements also increase security threat surfaces. With disaggregation, the core network can be implemented on geographically separated networks owned by different entities. Similarly, ORAN exposes additional threat surfaces. A study of gaps within ORAN concluded that insider attacks remain extremely dangerous even if all the optional security features are implemented. As Next G standards evolve, it is important to add options for mitigating insider attacks based on trust boundaries. For instance, if the gNodeB (gNB) base station is outside the trust boundary, it is important to have ciphering and integrity protection from the User Equipment (UE) to the UPF (which is not currently supported). Similarly, in a disaggregated network where network elements could reside on infrastructure owned by different parties, threat surfaces on the control plane also need to be investigated.
Security and performance: Since performance and Quality of Service (QoS) continue to remain a critical parameter for networks, providing security features with minimal impact on performance becomes key. Unfortunately, the performance optimization knobs on today’s cellular systems do not include security, but this issue can be addressed using a two-pronged approach: (1) Design fine-grain security knobs in next-generation networks so that security features can be used judiciously. (2) Develop a scheme to quantify the costs of providing different security features like TEEs, encryption, signatures, and other measures for a given system so the security features can be part of the performance and QoS design of the system.
It is also important to note that security policies need to have an architecture that it is consistent across the different elements of the network as well as consistently implemented across the different layers of the network (from the protocol layer to the hardware layer).
AI for security: AI algorithms can also be used to optimize the resources used by security mechanisms, especially in intrusion detection systems. Furthermore, when the packets are end-to-end encrypted, traditional tools like the open-source Snort network intrusion detection system cannot be used. Instead, AI can be leveraged to make assessments if there is an attack based on traffic patterns. In addition to intrusion detection, AI such as Large Language Models can also be used to detect gaps in a network’s security and develop code/solutions to address some of these gaps.
Security for the Future
It is known within the network security industry that with the emergence of quantum computers, existing asymmetric Public Key Infrastructure (PKI) encryption will be broken. Various Post-quantum Cryptography (PQC) techniques have been developed and are being standardized by the National Institute of Standards and Technology (NIST). To mitigate the quantum attacks, we need to replace digital signatures, key exchange, and public key encryption with post-quantum techniques. The 5G protocol uses the PKI in many places (in addition to the management and orchestration layer).
However, it is not a trivial task to replace standard algorithms with quantum-resistant algorithms. PQC algorithms have large keys, ciphertext, and signatures that could impose a significant communication overhead. Furthermore, since the payloads are much larger, it may take multiple transmissions to transmit the keys and signatures. Communication protocols will need to be modified to accommodate this issue, along with addressing state and integrity challenges. Latency, which is also critical in communication networks, will need to be addressed when using PQC techniques.