As the world tries to grapple with the implications of 5G, researchers from China have already started looking into 6G. 6G will operate on a much higher frequency than 5G, and the shorter wavelengths will allow for higher localisation accuracy, possibly down to centimetre level positioning.
A team of researchers from NTU Singapore, Monash University, Australia, Heilongjiang
University, China, and others investigated the challenges of embracing 6G as the world moves towards ML heavy solutions. Their main objective is to find out how to make ML more feasible in a high-speed wireless environment. Federated learning, stated the authors, is an emerging distributed AI approach with privacy preservation nature is particularly attractive for various wireless applications, especially to achieve ubiquitous AI in 6G.
Why Federated Learning
Traditional Machine Learning techniques rely on a central server and are prone to critical security challenges, e.g., a single point of failure.
Moreover, centralised data aggregation and processing cause large overhead and the researchers warn that the traditional centralised ML schemes might not be suitable for 6G.
This is where Federated Learning (FL) comes into the picture, which has become popular for its decentralised ML solution.
The choice of federated learning is mainly for two reasons: its distributed nature and privacy. Because in a 6G world, believe the authors, AI will bridge human-centric development with all aspects of network systems. Therefore, the security and privacy requirements of 6G communications are significantly higher.
Implementing Federated Learning
The procedure of FL-based architecture, note the authors, is divided into three phases:
In the first phase where the initialisation happens, a device will evaluate its service requests and decide whether to register with the nearest cloud for training an ML model via 6G. The cloud will also send initialised or pre-trained global models to each selected device.
The training phase deals with selecting each device and training a global model by using a local dataset to obtain the updated global model in each iteration.
Next comes the aggregation phase, where the cloud receives model updates of all selected devices for aggregation to obtain a new global model for the next iteration.
However, the authors also admitted that implementation is not straightforward as there are a few challenges: –
Expensive Communication: FL involves thousands of devices participating during model training, communication is a critical bottleneck for FL being widely used in 6G
Security & Privacy: the capabilities of each device in the network may vary with hardware (CPU or GPU), network connectivity (4G, 5G, 6G, WiFi), and energy (battery level). This heterogeneity, the authors believe, will bring the lead to flaws in the FL model and 6G network.
Although federated learning garnered its reputation for protecting the privacy of edge devices, procedures such as model updates (e.g., gradients information) instead of the raw data can still be disclosed. They can be prone to adversarial attacks such as membership inference or gradient leakage attacks.
For preserving privacy, the authors recommend techniques such as differential privacy, deep net pruning, and gradient compression.
Going Forward
Today, as the world tries to figure out a feasible 5G solution, 6G might look like a distant dream. But even to have something like 6G in the future, the researchers firmly believe in starting to look for solutions today because the challenges are immense.
One such speculation is that 6G communications can achieve up to 1 Tbps data rate per user with low latency and high end-to-end reliability.
Watch this video:https://www.youtube.com/watch?v=Z5vxRC8dMvs