The COVID-19 lockdown, along with other industries, has majorly disrupted the financial sector. Although the majority of the banking functions have already been digitised, onboarding customers has continued to be a time consuming in-person process. KYC or know your customer has been a critical and mandatory process for financial institutions; however, with a lockdown in place, all physical modes of submitting KYC applications have come to a standstill.
The traditional KYC process involves officially conducting valid document verification in-person (or virtual), which was not only consuming a lot of time and effort, and hampering the businesses, but also made the onboarding experience quite cumbersome for new entrees.
Thus, like any other agile organisation, Aditya Birla Sun Life AMC has also been working on various nuanced technologies to resolve this issue for their business. When asked about the challenges, A. Balasubramanian, the MD & CEO of Aditya Birla Sun Life AMC said, “The physical process has been quite cumbersome, and the onboarding experience of new investors was fraught with difficulties. As a result, the onboarding experience was acting as a barrier for entry.”
Therefore, Aditya Birla Sun Life AMC coupled their agility with Signzy’s technology — a fintech company — to respond to the logistical challenges thrown by the pandemic, enabling both business continuity along with safeguarding customers. The need of the hour called for Signzy’s AI-based Video KYC solution that would allow remote customer onboarding quickly and seamlessly.
Overview Of The Solution
Signzy’s AI-based Video KYC solution enabled real-time and remote customer onboarding, along with verification and fraud detection for Aditya Birla Sun Life AMC. The highly sophisticated solution offers 98% faster KYC process than the traditional systems and enables a zero-contact process that has been designed to replace the physical submission process of KYC documents.
Designed for banking grade technology, the AI-based Video KYC solution comes with the strictest infosec regulations and data security requirements. Further, the solution has matured over dialects, browsers and low-internet scenarios. The solutions work on real-time PAN verification, facial and image recognition technology to match the face of the customer, geolocation capture and IP check, along with end-to-end encryption and video forensics for spoof detection.
According to Arpit Ratan, co-founder of Signzy — the first step to the Optical Character Recognition pipeline was to create an algorithm to precisely crop out the region of interest to get the job done easier. For this, the company approached the deep learning technique to build a regression model which can predict the edges of the document that needs to be processed. After implementing a shallow custom architecture for predicting the outputs, the team managed to achieve good performance from the model.
Documentation rotation was a critical aspect of processing the documents; therefore, the company built a classification model which predicts the angle of the document. Post that, they also created an algorithm that can localise the text regions for further processing. For this, the team benchmarked various open-source text detection models on the test datasets.
Finally, the most critical step in the OCR engine is localising the text regions in the document, which has been done with the technique called — word-level classification. This method leverages deep learning, where the full text localised region is passed into an end to end pipeline to get the predicted text. Further, the cropped text region is passed into a convolutional neural network for spatial feature extraction and then passed on to the recurrent neural network for extracting temporal features.
Facial Recognition System For Detecting Customers
Explaining the process, Ratan, stated, to make this work, Signzy works with their in-house facial recognition system — Deep Auth that allows real-time authentication of large-scale using everyday devices like mobile phones, tablets, laptops etc. “The solution learns the facial features dynamically with online training with 150k iterations on the WIDER Face dataset, to recognise appearance variations like customers having a beard or wearing spectacles,” said Ratan.
The technology behind Signzy’s facial recognition system leveraged a series of convolution neural networks, which has been divided into two parts — face detection and face recognition.
Convolutional neural networks for face detection
Firstly, it involves three stages of convolutional neural networks, which ensure that the face of the customer is detected with accuracy. To facilitate this, the Signzy’s team first proposed regions of objectablility score and their regression boxes, which was then followed by utilising these regression boxes as the input. For this, the team applied non-maximal suppression to reduce the number of false positives. Following that, the team detects the facial landmarks with a five-point localisation to ensure that the faces of customers are correctly aligned. The loss function is the combination of the centre loss with Intersection Over Union loss.
Once the facial features are extracted, it is then passed to a siamese network for using the contrastive loss to converge the network. The team then uses K- Nearest Neighbours, which was injected during training, that classifies each encoding to the nearest face encodings. For facial recognition, the company used 512-D vectorisation, which helps in distinguishing the fine details of the customer’s face and provides high accuracy to the system.
Further to this, while evaluating the model, the company realised that the solution performs impressively on Face Detection Data Set and Benchmark. Here, the team used two images to train each of distinct faces and evaluated them with the test images — which highlighted a score of 99.5629 and 91.2835 accuracy respectively.
With its REST API interface, the solution is suitable for online training and real-time recognition of customers’ faces. Moreover, its advanced technology is also robust in detecting ageing and appearances, making it an ideal solution to deploy in remote areas.
With Signzy’s AI-based KYC solution, Aditya Birla Sun Life AMC was able to address the logistical challenges that the company was facing, along with easing their remote onboarding process. It not only witnessed a rapid increase of use on a daily basis by its partners and investors but also enhanced their customer experience, as a whole. Amid COVID, the company has witnessed a doubled up registrations which led to onboarding over half a lakh new investors to the mutual fund industry.