Duplicate Aadhaar is a big problem, open-source Bharat ABIS tool could help fix it

A biometric database is one of the most secure systems for verification. But at scale, where you may have data of over 1.5 billion people – in the case of Aadhaar – even this starts to be pushed to the limit. Researchers associated with the Unique Identification Authority of India (UIDAI) have come up with Bharat ABIS – an open-source biometric verification system that seems to be much better in speed and accuracy than the systems we use today.The research team included six UIDAI researchers, one researcher from IIIT Hyderabad and renowned biometrics researcher Anil Jain from Michigan State University. The study is titled, “Towards Billion-Scale Multi-modal Biometric Search.”The study examines how Bharat ABIS – an open-source multi-modal biometric system – handles large databases while still maintaining accuracy and speed. That is, how does this system manage to add new IDs into a large database without causing duplicates or wrongly flagging entries. To give you some context, in 2022, the Ministry of Electronics and Information Technology (MeitY) informed the Lok Sabha that nearly 6,00,000 duplicate Aadhaar cards had been identified and cancelled by May that year. Researchers say that ABIS was evaluated against a gallery of 220 million Aadhaar records. Though the system was trained on a smaller dataset, ranging between 1,00,000 to 2,00,000 entries. The data was de-identified before being fed to the system. What is Bharat ABIS?ABIS stands for Automated Biometric Identification System. Bharat ABIS combines fingerprints, iris, and facial data of an individual which is then processed into a single template that has a size of 13.5kb. This template is then matched against the database to find any pre-existing credentials – a pre-existing ID of the same person. To ensure that the system can handle entries at scale, ABIS comes with a scheduler that divides search work across servers. While the system can also use individual biometric data, such as only fingerprints or only iris, accuracy improves drastically when all three layers are combined. How good is Bharat ABIS?As per the study, Bharat ABIS has a False Positive Identification Rate (FPIR) of 0.1 per cent. FPIR refers to the error rate where the system may wrongly flag duplicates in the system. UIDAI’s target rate for vendors is 0.5 per cent – ABIS has a lower error rate.Bharat ABIS’ False Negative Identification Rate (FNIR), which refers to the error rate of missing duplicates is 0.05 per cent – the target set is 0.1 per cent. Do note that both the scores here apply for a 20 million database.Researchers compared Bharat ABIS to three pre-existing tools used inside Aadhaar. As per the study, ABIS achieves “comparable or better” accuracy than all three tools on the 20 million database. The system’s DeepPrint fingerprint model, when combined with vendor tools, can reduce the number of authentication failures by half, the study states. The biggest gains come for older age groups and manual labourers – who are likely to have worn fingerprints. This can help in millions of transactions everyday.At the same time, the study says Bharat ABIS can make 100 searches per second on a gallery of 40 million records in a single server that uses 8 Nvidia H100 GPUs with 2TB RAM. The researchers add that ABIS could be the go-to biometric identification in about 160 countries that have a population of under 20 million. These countries, the researchers say, may not be able to afford off-the-shelf systems from vendors, but could adopt ABIS thanks to it being open-source.However, the study does mention that ABIS still needs to be tested at scale before we can draw a conclusion on whether it can truly work with large databases such as India’s Aadhaar system.That is to say that while Bharat ABIS seems to be a step in the right direction in preventing duplication of IDs like Aadhaar, it may be too early to tell whether we can see such an open-source platform be used just yet.- Ends