Application of Machine Learning Algorithms in Audit By Mr. Deepak Raghu, SrDAG

Machine mimicking the mind

There has been a tremendous explosion of data available with public sector entities over the last 20 years. This is largely driven by higher deployment of digital devices such as sensor instruments, scanners, meters, smart phones, CCTV cameras, satellites, and drones as well as software applications such as Supervisory Control and Data Acquisition Systems, Geographic Information Systems and biometric ID based applications for citizens.

The resulting granular, variegated and voluminous data has chanced upon significant reduction in the costs of storage and processing of data. These conducive circumstances have resulted in the emergence of Machine Learning (ML) Algorithms. These are computer algorithms, which mimic the techniques used by the human brain to identify patterns, learn to create mental models and solve problems. These algorithms use a high volume of input data to train and learn, to create an output data model. This output data model is then used on test data to deliver the required results.

For example, one can envisage an ML algorithm, which is intended to work as an image classifier for defects in construction i.e., the ML is required to deliver the result of whether an image contains a construction defect such as a crack in concrete or a bent iron beam or not. In order to deliver this result, the ML algorithm would have to use a high volume of construction images, which are clearly labelled as either defective or defect-free, as input data to train and learn to identify the different types of defects. This learning would then result in creation of an output data model. The output data model would then be in a position to consume new images, the test data, which are not labelled and classify them as either defective or defect-free. The accuracy of the output data model would depend on the volume and quality of the input training data.

Use of machine language in audit

Such ML algorithms would find ready-use cases in Government Audit. For the high volume of structured data being maintained by audited entities, the application of analytical tools such as IDEA and SQL would suffice. But for the combination of high volume of unstructured data such as images and videos of documents and objects and structured data being maintained, the application of ML algorithms in Audit would appear to be the logical solution.

ML algorithms may be applied to solve the following types of basic problems-

ML algorithms enable detection of sophisticated cases of non-compliant actions / potential frauds from the test data,even though it may not be clear to the Auditors a priori, as to how the data fields may be internally correlated.

In terms of the types of Audit assignments in which ML algorithms could potentially be applied, the following are some illustrative examples where these algorithms can be applied for Image Analytics, with the images obtained from satellites, drones or CCTV cameras-

The use of ML algorithms during Audit is poised at a very interesting stage, with a large scope for innovation and delivery of insights and results that would not be feasible through human analysis alone. It is for this reason that SAIs around the world are deeply engaged in identifying and developing strong uses for ML algorithms.