AI/ML security in mobile telecommunication networks
- AI/ML technologies offer promising advancements in several areas including security, but they also amplify the effectiveness and reach of cyber threat actors’ offensive capabilities.
- AI/ML technologies open new frontiers in cybersecurity defense.
- It is crucial that AI/ML-driven systems embedded within the mobile network are robust and secured.
- The primary policy goal should be to continue to safeguard users of mobile networks, ensuring that AI/ML components used in networks align with security and privacy standards and requirements.
- Policymakers should enable the industry to maximize benefits from AI/ML, including in the security domain.
From a mobile telecommunication network security-centric perspective, Artificial Intelligence (AI) and Machine Learning (ML) can be approached in the following ways:
- AI/ML as tools employed by threat actors to attack mobile telecommunication networks,
- AI/ML as tools to enhance mobile telecommunication network security, and
- AI/ML integrated into mobile telecommunication networks as attack targets, necessitating robust defense strategies.
When adopting new technologies like AI/ML, there is always a trade-off between benefits and risks. It is essential to recognize that risks vary across different deployments, requiring the selection of appropriate security controls tailored to each specific context. Therefore, a security risk assessment of AI/ML in context of mobile telecommunication networks is necessary to understand how current state-of-the-art security measures stand up. Hereby developing an understanding to ensure that network products and services continue to offer the best-in-class security.
The primary policy goal should be to continue to safeguard users of mobile networks, ensuring that AI/ML components used in these networks align with security and privacy standards and requirements. Policymakers fostering collaboration between the public and the private sector stakeholders is key to ensuring that net benefits from AI/ML are maximized. A synergistic partnership between government and industry is the most promising path to bolster cyber defenses to ensure the security of subscribers and the resilience of mobile networks.
AI/ML as attack tools
AI/ML technologies offer promising advancements in security with its efficiency and unique automation capabilities, but they also amplify the effectiveness and reach of cyber threat actors’ offensive capabilities. This challenge requires coordinated actions. Hence, responsible stakeholders must consider the impact of unlawful use of AI/ML on cybersecurity risks since such use of AI/ML technologies can substantially change both their type and scale. While AI/ML might not always introduce new types of attacks by themselves, they amplify the effectiveness and reach of existing attack methodologies, such as enhancing attack automation, identifying the most vulnerable or valuable targets and adapting attacks more responsively and intelligently.
AI/ML as tools to enhance mobile telecommunication network security
While AI/ML can be leveraged as an attack instrument to disrupt the robustness, there are ways to use AI/ML on the other hand to improve security across mobile telecommunication networks. AI/ML technologies open new frontiers in cybersecurity defense. AI/ML can enhance threat detection in mobile telecommunication networks by supporting traditional methods and identifying potential new threats. AI/ML overcomes the limitations of traditional signature-based detection techniques, can identify new or complex threats, and apply more dynamic and adaptable security controls. Specialized AI-based security controls can employ advanced behavioral analysis and real-time adaptation, helping match the evolution of attacker techniques.
Also, the rise of offensive AI requires advanced countermeasures. Specialized AI-powered security controls, such as AI-driven intrusion detection systems, can serve as advanced mechanisms, designed to detect AI-driven attacks by identifying AI-specific anomalies. Real time adaptation is particularly important against AI-based attacks, which can adapt themselves faster than humans can react. For example, if an AI-driven malware changes its signature to evade detection, a specialized security system can update its algorithms almost instantly to recognize this newly evolved threat.
Securing AI/ML components in mobile telecommunication networks
Mobile telecommunication networks serve as the backbone for transmitting voice and data across the globe, enabling seamless connectivity for user devices like mobile phones. These networks are structured into five main logical parts (see Figure 1): The Radio Access Network, core network, transport network, management, and interconnect network. Those parts operate on three distinct planes—control, user, and management—responsible for signaling, payload, and network management traffic.
Security of AI/ML should focus on the following components containing AI/ML technologies: Next-Generation Radio Access Networks (NG-RAN, e.g., 5G RAN), access networks based on the O-RAN architecture, NWDAF within 5G core network, AI-driven Operations/Business Support Systems (OSS/BSS), and security management tools. Integration of AI/ML there without appropriate security measures might elevate security risks in mobile telecommunication networks.
Even though AI/ML technologies bring a unique set of threats, they are not standalone units but are integrated into traditional mobile telecommunication systems. A comprehensive threat analysis should thoroughly assess the entire system, considering both threats to non-AI/ML entities and the environment in which AI/ML components run, as well as AI/ML-specific threats. This requires the development of defensive strategies that encompass both traditional and AI-specific security controls to maintain a robust security posture against a broad range of potential attacks.
AI/ML development and environment threats target vulnerabilities in components surrounding the AI/ML, such as the execution environment or data storage, rather than the model itself. These can be traditional forms of attacks, not specifically designed to exploit AI/ML vulnerabilities, but can still compromise the AI/ML system, especially if it automates existing processes. An attacker exploiting traditional vulnerabilities could even unknowingly compromise the AI/ML system, as well as the mobile telecommunication function it supports. Such attacks might also target the software development life cycle, deployment procedures, or communications and can result in compromised components or new exploitable vulnerabilities.
AI/ML-specific threats target AI/ML models and the associated data within the mobile telecommunication infrastructure. Adversaries may seek to corrupt or tamper with these models for various malicious objectives, from disrupting service to intellectual property theft. According to NIST AI 100-2 E2023, such threats include evasion, poisoning, and privacy attacks. Evasion attacks seek to modify the behavior of AI/ML models through specially crafted inputs, including prompt injection attacks against generative large language models (LLMs). Poisoning attacks contaminate training datasets or model parameters to insert backdoors or reduce model performance. Privacy attacks focus on extracting insights from the training data—whether targeting the data itself (data privacy attacks) or the underlying ML model (model privacy attacks). This includes scenarios where generative LLMs memorize and output training data verbatim.
AI/ML threat mitigation in mobile telecommunication networks
To mitigate attacks targeting the AI/ML systems, the first step is to identify threats to the AI/ML environment and AI/ML assets by performing a comprehensive security risk assessment. Appropriate security controls, including both traditional and AI/ML-specific measures, are essential. The risk assessment, across the entire AI/ML development and operational lifecycle, helps identify and prioritize the necessary security controls.
When there is a need to mitigate an identified threat, the first layer of defense involves implementing traditional security and privacy controls. These measures are effective against well-known threats and can also address some AI/ML-specific threats.
In addition to traditional controls, AI/ML systems might require specialized security controls to protect against unique types of attacks like evasion. Techniques for detecting and preventing AI/ML-specific attacks are outlined in NIST AI 100-2 E2023 or OWASP ML Security Top 10. Since AI/ML attacks and mitigations are subject to ongoing research, the suitability of these methods requires further investigation and should be tailored to specific use cases.
Ericsson’s holistic security approach depicted in Figure 2 is founded on the Ericsson trust stack and designed to address and manage security challenges effectively. The four layers in Ericsson’s trust stack have been adapted to secure AI/ML components within mobile telecommunication networks. This adaptation incorporates findings and recommendations from academic research, industry insights, and cybersecurity authorities’ knowledge about AI/ML risks and mitigations, as detailed in guidelines from respected bodies like the NCSC and CISA.
Security policy recommendations to policymakers
As AI/ML technologies become integral to mobile telecommunication systems, the complexity of cyber-attacks and defenses is expected to rise. Security measures must be flexible, adaptive, and scalable as AI/ML models grow in complexity and data volume. Important activities include:
- Developing standardized secure protocols and frameworks for AI/ML.
- Collaborative efforts in academia, industry, and standards organizations to address AI/ML's unique security challenges, such as standardized risk management, security assurance, and governance tools, processes, and methodologies.
- Open-source projects, tools, and shared research that contribute significantly to advancing AI/ML security.
Mobile networks, as part of national infrastructures, are already subject to comprehensive regulations for security and privacy, which consequently will also influence the security requirements of AI/ML components used in mobile telecommunication networks. In addition, upcoming AI/ML-specific regulations will set requirements that will ensure the security of mobile telecommunication networks. The primary policy goal should be to continue to safeguard users of mobile networks, ensuring that AI/ML components used in networks align with security and privacy standards and requirements.
Furthermore, policymakers should enable the industry to maximize benefits from AI/ML, including in the security domain, but do so in ways that do not slow innovation through over-regulating while avoiding compromising targeted security and resilience objectives. This can be achieved by considering the following recommended cybersecurity policy actions:
- Stimulate research on AI/ML to enhance threat detection.
- Engage with industry to share understanding of telecom-specific threat landscape, focusing on 5G evolution and future 6G systems.
- Promote awareness of structured approaches to AI/ML-specific attacks.
- Work with industry to develop sector-specific security best practices for mitigating AI/ML-enabled traditional and novel attacks.
- Foster broad implementation of relevant security controls.
- Update security risk assessments in existing regulatory frameworks and include relevant AI/ML risk assessment practices, leverage NIST AI RMF.
- Promote coordinated standardization work, while avoiding creating silos.
- Promote AI/ML security best practices where appropriate supported by public funding of open-source projects.
- Promote advancements in confidential computing to enable Trusted Execution Environments for virtualized AI/ML training and applications.
- Promote privacy-enhancing technologies to protect privacy remove barriers to data sharing without violating privacy.
Sign up for the 5G Security blog series
Don't miss out - sign up today and be notified of each episode as it is released.
Sign up nowAs we navigate the intricacies of the digital age, one fact remains resolute: the security of these critical networks is not a one-time effort but an ongoing imperative. It’s also not a one-person or organization job. It requires vigilance, innovation, and partnering with trusted experts to help you find, face and defeat the threats at hand. It needs a collective commitment from all of us in the telecom industry, and adjacent ecosystems, to work together and safeguard our modern communication ecosystem.
Learn more
MLSecOps: Protecting the AI/ML Lifecycle in telecom white paper.
Trustworthy AI - What it means for telecom white paper.
Four benefits of AI for security, safety and transparency in telecom blog.
Find out more about telecom security for a connected world and about Cyber network security in the era of 5G - Ericsson
Learn more about staying ahead in this evolving landscape in our 5G security blog series.
Explore telecom AI
Explore AI in networks
RELATED CONTENT
Like what you’re reading? Please sign up for email updates on your favorite topics.
Subscribe nowAt the Ericsson Blog, we provide insight to make complex ideas on technology, innovation and business simple.