Wireless network operators are preparing for a generation of infrastructure that will have AI built into the architecture from the beginning. Sixth generation networks, expected to reach commercial development over the next decade, are being designed with AI at the heart of how they allocate spectrum, route traffic, and detect failures.
A paper by researchers at Harokopio University in Athens examines how different AI techniques map to specific layers of 6G network design, from the physical radio layer to network management and service delivery. The paper covers publications from 2018 to 2025 and is based on the standardization efforts of 3GPP, ITU-T focus group on 6G, and O-RAN ALLIANCE.
What to expect from 6G
The performance goals for 6G are significantly higher than those for 5G. It is predicted that data transfer speeds could exceed 10 terabits per second, compared to about 10 gigabits per second in current 5G deployments. End-to-end latency is targeted at approximately 0.1ms, a 10x improvement over the 5G requirement of 1ms. Reliability targets for ultra-critical applications reach 99.9999 percent, and the specification covers use cases such as autonomous vehicle control, remote surgery, and industrial automation.
Coverage is also expected to extend to deep sea, underground, and space environments, supporting connectivity in locations that current networks cannot reach.
How AI is partitioned across the network stack
Researchers organize AI techniques based on where they operate within the network. Traditional machine learning techniques are applied at the physical layer to handle tasks such as channel estimation and beam optimization, including working with reconfigurable intelligent surfaces. Deep learning and reinforcement learning operate at the network and management layers to support spectrum allocation, network slicing, and real-time orchestration.
Federated learning is primarily assigned to the service layer and allows devices to train shared models without sending raw data to a central server. This approach is relevant for IoT deployments, healthcare applications, and augmented reality services where data sensitivity or bandwidth constraints make centralized training impractical.
Explainable AI works across all layers to address the need for transparency in automated decision-making. This is a requirement in line with regulations including the EU’s GDPR.
Security concerns associated with AI deployment
Integrating AI into 6G also introduces security risks not present in traditional network architectures. AI systems trained on large datasets can be targeted by data poisoning attacks, where malicious input degrades model performance. Despite its privacy benefits, federated learning remains vulnerable to model inversion attacks that can extract information from shared model updates.
Generative adversarial networks can generate synthetic network traffic and fake credentials that bypass traditional intrusion detection systems.
Countermeasures being considered include adversarial training, Byzantine fault-tolerant aggregation in federated systems, and AI-driven anomaly detection that monitors traffic patterns in real time.
Blockchain infrastructure is being valued as a support layer for audit trails and identity management in decentralized AI deployments, as well as lighter-weight consensus mechanisms such as Proof of Stake that have been proposed to keep energy costs manageable.
Energy and hardware constraints
Running AI at scale within a network incurs energy costs. The computational burden of training and running large models across dense device deployments conflicts with the sustainability goals that 6G carriers are expected to achieve. Research directions include model compression, quantization, and pruning to reduce the computational burden of AI inference without significantly compromising accuracy.
Hardware development is also a constraint. Terahertz communications, which operate in the 0.1-10 THz range and are central to 6G’s high-speed targets, require new transceiver designs and face significant path loss challenges. Edge computing microchips that can perform AI inference locally are necessary for latency-sensitive applications, but current chip designs have yet to resolve the tension between processing power and power consumption.
Quantum computing is emerging as a longer-range possibility. Quantum algorithms, such as quantum approximate optimization algorithms, have the potential to address resource allocation problems at scales beyond the capabilities of classical optimization. Integration with existing infrastructure remains an engineering challenge, given current requirements for operating temperatures near absolute zero and error correction overhead.
Researchers note that interoperability between vendors and regions remains an open issue. For AI components from different manufacturers to work together within a single network deployment, standardized APIs and data exchange formats are required.
Download: Tines Voice of Security 2026 Report
