Large language models: GenAI that works for India's defence sector

Security agencies need GenAI for intelligence gathering, logistics and autonomous combat systems

Generative artificial intelligence (GenAI) has taken every possible business, industry and government department by storm. India’s defence sector is not far behind. The rise of large language models (LLM), which are used to power GenAI products, has not left defence forces untouched.

India’s Ministry of Defence set up an AI taskforce in 2018. By 2022, around 40 different GenAI products developed by defence public sector enterprises were in place.

The application of AI-based technology in defence covers many functions and possibilities such as training, surveillance, logistics, cybersecurity, unmanned aerial vehicles and others.

The defence and security sectors use LLMs for automating data processing, converting information to intelligence, analytics, summarisation and other purposes. (LLMs, according to one definition, are AI systems capable of understanding and generating human language by processing vast amounts of text data.)

One security aspect to flag is that many defence institutions use the services of companies that provide localised and disconnected LLMs trained on existing sector-specific datasets and not connected to any Cloud service real time. 

An offline LLM is less likely to suffer unauthorised access and distributed denial-of-service (DDoS) attacks, reducing the risk of data leakage. In a DDoS attack, multiple computers overwhelm a website or online service with excessive traffic and make it unavailable to users.

Safe offline

LLM integration with intelligence systems can reduce the time and effort of an organisation from days to just 15-30 minutes, said Tarun Wig, chief executive officer and co-founder of Innefu Labs. The Delhi-based AI security solutions company provides offline LLM infrastructure and security solutions to the Indian Army and the Border Security Force. 

“There are no limitations at all to LLM being offline. They can be (and are currently in use by security and surveillance agencies) for profiling individuals and organisations, generating reports on events. They can also be connected to verifiable OSINT sources to generate reports on real time trends and events,” said Wig. OSINT, short for open-source intelligence, involves gathering and analysing publicly available information from various sources for security purposes.

“Offline LLMs are desirable for security establishments. You don’t want such LLMs to be exposed to the external world from certain security perspectives, assuming that such LLMs will be fine tuned with internal data that can’t be shared with others who don’t have appropriate access,” said Naresh Singh, senior director analyst, Gartner.

Singh said LLMs for the defence sector can leverage the library of information (structured and unstructured like images and videos) already learned from the internet and other online repositories.

“Once trained, these LLMs can be effectively hosted offline and only retrained when necessary at intervals to update with new data. It is not required for LLMs to be connected to the internet in real time, always,” said Singh. 

“Of course, being connected always has its own benefits in terms of reinforced training of new data available through live feeds of information but the LLM could still be useful without being connected live to the internet,” he said.

Localised edge computing – a distributed computing framework that enables data to be processed closer to where it is created – will be crucial for offline implementation of Gen AI technologies.

"India's defence and security establishments stand to benefit significantly from integrating offline LLMs and with distributed edge computing centres. Processing vast amounts of data in real time is crucial for operations such as surveillance, logistics, and autonomous combat systems,” said Som Satsangi, senior vice-president and managing director, Hewlett Packard Enterprise India.

Data security

For an individual, not being connected to real-time internet is a disadvantage but for defence and security it is a need. 

Wig anticipated that security establishments will expand the use of LLMs in the next few years, but a major hurdle will be the government procuring sub-standard services from public sector undertakings.

“India seems to be going back towards the licence raj system where some PSUs (public sector undertakings) and certain newly formed academic institutions are being used to sell below average products at exorbitant prices after white labelling them,” said Wig, referring to a business practice in which a company buys its product from another and rebrands it as their own.

“I think the government needs to evaluate the technology on its merit and decide to go with the best.” 

LLMs are likely to carry biases and in the case of critical and sensitive sectors it could pose risks. 

“Biases creeping in are a challenge that any government and public use of LLMs have to address. The government must come up with regulations to this effect,” said Singh, the Gartner executive. 

He suggested having sovereign LLMs for security and defence sectors. “There is no better alternative than having our own models that we can really trust upon.”