Page 50 - Banking Finance August 2023

FEATURES

CAG focuses on AI to uncover fake

claims, conduct performance audits

rom identifying non-existent schools claiming “The cases detected helped in identifying the risky
F scholarship benefits to detecting similar images transactions, duplicate beneficiaries, fake and ineligible
beneficiaries,” it stated.
used by multiple beneficiaries of government
schemes, the Comptroller and Auditor General of
However, these findings are subject to field validation, and
India (CAG) is using Artificial Intelligence (AI) and Machine
Learning (ML) in a big way. the actuals will be reported after field audit, a source said,
adding that the audit report will be out soon. “These are
techniques being followed in phase-1 data-driven audits,
The CAG has showcased at least half-a-dozen such case
using which phase-II field validations will be done by using
studies showing use of these tools in a ‘Compendium on
risk-based sampling, after which audit reports are finalised,”
Responsible Artificial Intelligence’, unveiled at the SAI20
the source said.
Engagement Group Summit held under India’s G20
Presidency at Panaji recently.
The CAG also used an AI & ML model to detect non-existent
schools claiming scholarship benefits. “Machine learning
One of these case studies is about the use of AI in detecting
model was developed in Python to identify suspected fake
duplicate, fake and ineligible beneficiaries of Digital
schools which claims scholarship in 2017-18, based on pre-
Saksharta Abhiyan (DISHA), the government’s digital literacy
defined risk parameters identified from the data pertaining
programme. As per the scheme’s guidelines, the beneficiaries
to 2019-20. A total of 17 parameters at school/institute
trained under the programme were required to upload their
level were identified for the model. A set of 10 different
photographs.
Machine Learning Algorithms was used with two different
techniques…,” it stated. “The model achieved above 92%
To see whether same images or different images of the
accuracy. This helped in identifying the risky samples for the
same beneficiaries, or non-human images, were used for
field level verification.”
claiming the training cost, the centre for data management
& analytics at CAG, which does research and analysis for
phase-1 data-driven audits, developed an intelligent model
using open-source platform — Python — to automatically
analyse these images. With the help of this model, the large
volume of images was analysed automatically, which was
not possible manually.

The compendium mentions that two images of a person
were used 316 times, while two images of another person
were used 187 times. Similarly, eight photographs of a
person were used 78 times and non-human images were
found in several cases, it shows.

50 | 2023 | AUGUST | BANKING FINANCE