New Delhi: Researchers associated with the Unique Identification Authority of India (UIDAI) have developed Bharat ABIS, an open-source biometric identification system designed to detect duplicate identities at very large scale.

The system combines fingerprint, iris and facial biometrics and has been evaluated against a gallery containing 220 million Aadhaar records, according to a research paper titled “Towards Billion-Scale Multi-modal Biometric Search”.

The researchers say Bharat ABIS can deliver accuracy and search performance that are comparable with, or better than, existing biometric identification systems used in Aadhaar-related applications.

However, the study also makes clear that further testing at much larger scale is required before conclusions can be drawn about deploying the system across the entire Aadhaar database.

Why duplicate Aadhaar records are a concern

Aadhaar is one of the world’s largest biometric identification systems, with the database operating at a scale that creates significant technical challenges for identification and de-duplication.

In 2022, the Ministry of Electronics and Information Technology informed the Lok Sabha that nearly 6 lakh duplicate Aadhaar numbers had been identified and cancelled by May that year.

Biometric identification is intended to prevent the same person from obtaining multiple identities. However, matching a new enrolment against a massive database requires systems that can process enormous numbers of biometric comparisons without generating excessive false matches or missing genuine duplicates.

This is the problem Bharat ABIS is designed to address.

The research team included six UIDAI researchers, a researcher from the International Institute of Information Technology, Hyderabad, and biometrics researcher Anil Jain of Michigan State University.

What is Bharat ABIS?

ABIS stands for Automated Biometric Identification System.

Bharat ABIS is described as an open-source, multi-modal biometric identification platform. Rather than relying on just one biometric characteristic, the system combines fingerprints, iris information and facial data.

The biometric information is processed into a single template that is approximately 13.5KB in size.

When a new identity is added, the template can be compared with existing records to determine whether the same person may already have an identity in the database.

The system can also perform searches using individual biometric modalities, such as fingerprints or iris scans. However, according to the study, combining all three biometric modalities substantially improves identification accuracy.

How Bharat ABIS handles large databases

Processing hundreds of millions or potentially billions of biometric records requires more than simply having an accurate matching algorithm.

Bharat ABIS includes a scheduler designed to distribute search workloads across multiple servers. This allows biometric searches to be divided into smaller tasks and processed in parallel.

The researchers evaluated the system against a gallery of 220 million Aadhaar records. The data used for the evaluation was de-identified.

Although the evaluation involved the much larger gallery, the system was trained using smaller datasets containing between 1 lakh and 2 lakh entries.

This distinction is important because performance in a controlled research environment does not automatically establish that a system will perform in exactly the same way when deployed across a database containing more than a billion records.

Bharat ABIS records low error rates

The research paper reports two important measures of biometric identification performance: False Positive Identification Rate (FPIR) and False Negative Identification Rate (FNIR).

FPIR measures situations where the system incorrectly identifies someone as a duplicate. According to the study, Bharat ABIS recorded an FPIR of 0.1% on a 20-million-record database.

The study compares this with a UIDAI target rate of 0.5% for vendors.

FNIR measures cases where the system fails to identify an actual duplicate. Bharat ABIS recorded an FNIR of 0.05% on the same 20-million-record database, compared with a stated target of 0.1%.

Lower values for both measures indicate fewer false matches and fewer missed duplicate identities.

The researchers also compared Bharat ABIS with three existing biometric tools used within the Aadhaar ecosystem. On the 20-million-record database, the study reported that Bharat ABIS achieved “comparable or better” accuracy than the three systems.

Fingerprint technology could help with worn fingerprints

Another potential application involves improving biometric authentication for people whose fingerprints are difficult to capture.

The study says Bharat ABIS uses a DeepPrint fingerprint model that, when combined with existing vendor tools, can reduce authentication failures by around half.

The researchers reported particularly significant improvements among older people and manual labourers, whose fingerprints can become worn or less distinct over time.

This could have implications for biometric authentication in large-scale public and financial transactions, where even a small improvement in authentication reliability could affect a large number of users.

The research therefore looks beyond simply identifying duplicate Aadhaar records and examines how improved biometric models could reduce authentication failures.

How fast is Bharat ABIS?

The study also evaluates the system’s ability to conduct biometric searches at high speed.

According to the researchers, Bharat ABIS can perform about 100 searches per second against a gallery of 4 crore records using a single server equipped with eight Nvidia H100 GPUs and 2TB of RAM.

The researchers suggest that the technology could also be useful in countries with populations below 2 crore.

Because Bharat ABIS is open-source, countries that may not have the resources to purchase proprietary biometric identification systems could potentially adapt the technology for their own identity programmes.

The researchers suggest that the system could therefore have applications beyond India, particularly in countries looking for lower-cost biometric identification infrastructure.

Why Bharat ABIS is not ready to prove the Aadhaar-scale case yet

Despite the promising results, the research does not establish that Bharat ABIS is ready to operate across the full Aadhaar database.

The researchers themselves note that additional testing at larger scale is necessary.

The difference between evaluating a system on a 20-million-record database and deploying it across a database containing more than a billion identities is substantial. Search performance, infrastructure requirements, error rates and operational reliability can change as databases become larger.

The 220-million-record evaluation is therefore significant, but it should not be interpreted as proof that the system has already solved biometric de-duplication at Aadhaar scale.

There are also broader considerations around the security, governance and privacy of biometric databases. An open-source system can make technology more accessible and allow researchers to examine and improve the underlying software, but deployment in a national identification system would still require extensive security testing and operational safeguards.

Open-source approach could broaden biometric access

Bharat ABIS represents an attempt to develop biometric identification technology that can operate at very large scale without relying entirely on proprietary systems.

Its combination of fingerprint, iris and facial recognition, distributed search capabilities and reported accuracy results makes it a significant research development in large-scale biometric identification.

The reported performance against hundreds of millions of de-identified records indicates that the system can handle a substantial database, while the study’s comparison with existing tools provides an indication of its potential.

However, the researchers’ own qualification about further large-scale testing remains important.

For Aadhaar, where the consequences of both false matches and missed duplicates can be significant, performance at billion-scale deployment would need to be demonstrated rather than assumed.

Bharat ABIS could therefore represent a step towards more accessible and scalable biometric identification, but whether it can ultimately serve as a practical solution for duplicate Aadhaar detection will depend on further testing, validation and real-world deployment.