Page 240 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Key Aspects of Solu- 1) Firstly, a common data set, consisting of regional images and loca-
tion tion-based demographic information, to be created for transmitter
side and receiver side containing regional
2) Then both the descriptor (at the Tx side) and visualizer (at the Rx side)
LLMs are trained and fine tuned
3) Afterwards the descriptor module creates description of the visuals
which is transmitted over the GEO Sat link.
4) The visualizer module consumes the receive description and gener-
ates an image/ video frames semantically matching to the source
visuals.

Technology Keywords Generative AI, Large-language models, Semantic communication
Data Availability Private

Metadata (Type of Images, Location tag (optional, but desired)
Data)

Model Training and LLaMA model (version 3 released on Apr'24, version 3.1 released on
Fine-Tuning July'24, version 3.2 released on Sept'24) model is used for fine-tuning
using custom dataset.
Testbeds or Pilot The concept is yet to be fully deployed. The concept has been published
Deployments in [2].

2 Use Case Description

2�1 Description

Let us consider a scenario of a massive disaster (like flood) over a vast area, the terrestrial cellular
connectivity may get destroyed. This jeopardizes the ability of remote instantaneous visual
tracking of the disaster site by the central control room. An alternative might be that a local first
responded team transmits the visuals from the disaster sight using geostationary (GEO) satellite
backhaul from the rescue boat. Although Low Earth Orbit (LEO) satellites are promising for
providing broad-band internet service for future, but back-haul on LEO has several privacy and
security issues due to their vicinity to earth and mobility. This also gives rise to many challenges
due to the handover requirements and inter-satellite link congestion. GEO satellites continue
to be a reliable backhaul for critical utility services utilizing non-terrestrial network (NTN) .

Due to the huge latency and lack of bandwidth over GEO based backhaul, sharing high-
definition video and images in real-time may not be practically possible. As an alternative, we
can conceive the Semantic Communication based futuristic disaster tracking system as in Fig.
1. In this case a private network is set up with non-terrestrial backbone. The network access
point is installed on the floating large boat acting as the base unit of the recue party. The base is
equipped with an antenna connected to the control unit through GEO. The rescue party roams
around in a small boat equipped with user equipment (UE) which captures the live visuals and
transmits to the base unit edge computing service. The semantic encoder residing in the edge
computer extracts the relevant scene understanding and creates a semantic description against
each visual. The encoded semantic information demands much less bandwidth compared to
transmission of the full video or the image and can be transmitted over the GEO backhaul much

204