The Delhi Police is in the process of equipping its beat cops and Police Control Room (PCR) vehicles with remote facial recognition systems. This way, both local cops and first responders will be able to scan people’s faces using remote devices and run them against Delhi Police’s databases, DCP (Operations and Communications) S.K. Singh told MediaNama.

This won’t be Delhi Police’s first tryst with widespread use of facial recognition technology. It had earlier used the technology to screen crowds at Prime Minister Narendra Modi’s December 2019 rally in Delhi. Union Home Minister Amit Shah had declared in Lok Sabha that the Delhi Police had used its facial recognition system to trace 1,100 people who had caused riots in Delhi in February.

Delhi Police’s e-beat book app, used by beat cops, has already been equipped with facial recognition software and has led to a few arrests (more on that below).

Singh, who is charge of the 112 helpline, said that the Delhi Police is looking to use the Mobile Data Terminal (MDT), that all PCR vehicles are equipped with, to scan suspects’ fingerprints and faces on the field. The 8-inch HP tablet is already equipped with fingerprint verification to take PCR personnel’s attendance.

“Suppose you find a suspect, just take his finger, send it to the system, and it will tell whether he is registered in a case or not,” Singh explained. The Delhi Police wants to build on top of that and arm it with facial recognition technology (FRT) for suspects on the street. “MDT doesn’t have facial recognition presently but in the future we will do,” Singh said.

“We have just piloted the project. We have tested it. We have not yet started it,” Singh said, referring to both fingerprint verification and facial recognition.

However, even in the pilot stage, the process for implementing facial recognition has not been smooth sailing. “The plan is there but we are facing some problem[s] there because facial recognition requires huge bandwidth that we are not getting from 4G. So that is the issue. In case 5G is coming, then only we will be able to successfully implement the facial recognition. The limitation is the network,” Singh told us.

Since the project is still in pilot stage — Phase I —, and has not been rolled out yet, the standard operating procedures are not in place. “We just did a pilot project just to ensure that the system is working. It [protocol and SOPs] will be in Phase II. … Phase II proposal has gone to the MHA and once it is [approved], then it will be [implemented],” he said. Phase I started on September 25, 2019. “In that, testing and all we did but the whole proposal will be conceptualised in Phase II,” Singh said.

In terms of privacy protocols, Singh said that the photograph, if it is not of the suspect, is immediately deleted and removed from the directory. He was not sure if it meant that the log of the query itself, which could retain a copy of an innocent person’s face, would also be deleted from the central database. “[In the] central database, it will be known whose photograph was queried so that is NCRB [National Crime Records Bureau] but I don’t know whether they maintain or not,” he said.

““Data is still not integrated. We are working on that. We have already purchased everything. System is installed and in place but only the exchange of data is required because we are waiting for NCRB to open it [the centralised database that will underpin the planned Automated Facial Recognition System].” — S.K. Singh, DCP (Operations and Communications), Delhi Police

Currently, the Emergency Response Support System (ERSS), which runs the 112 helpline, has equipped all PCR vans with the MDT to which 112 call dispatchers send the location of a caller who might be in distress. The 112 app that runs on the MDT is developed by the Centre for Development of Advanced Computing (C-DAC), the Ministry of Electronics and Information Technology’s research and development organisation.

Beat cops already have access to the tech on mobile devices

Delhi Police’s e-beat book system, which is an app for beat cops, is already equipped with facial recognition technology that has been used to apprehend a pickpocket.

The app was developed by New Delhi-based IT firm Mobineers Info Systems and makes it possible for beat cops to fill out 114 forms related to their beat via a six-inch Samsung smartphone supplied by the Delhi Police, the company’s CEO Abhay Patodia told MediaNama. It has around 7,500-8,000 users — 1,752 beats that have 3-4 users each, around 900 division officers and 178 police stations.

“Facial recognition technology was already available with Delhi Police but that was available till police stations. Now we have integrated that technology also and made that technology available to all field constables, all beat staff,” Patodia said. It is understood that the app, which cost the Delhi Police ₹2 crore to procure as per Patodia, uses FRT that was supplied by another Delhi-based firm, Innefu Labs.

Thus, the device is supplied by Delhi Police, the app by Mobineers and the FRT by Innefu Labs.

Which database: Patodia said that the app is linked to the dossier database that is maintained by the Delhi Police but the Police can link it to whichever database it wants to. “We don’t have access to that database,” he said. “Server is by Delhi Police, device is by Delhi Police, user is Delhi Police. We have only given the software. … We don’t maintain a copy of it. We don’t save, we don’t delete it,” Patodia said.

How the app came to be: The tender for the e-beat book system was released in June and Mobineers rolled out the app in its first police district — South-West — on August 15. “Now it is rolled out completely in 15 [Delhi Police] districts,” Patodia told us. Two Delhi Police departments — Modernisation and Crime — were involved in the project.

What does the app do: “This app is unifying all the different datasets of Delhi Police and making them available in a common, unified format to the field constable,” Patodia said. A beat book maintains data of the local area, such as on banks, ATMs, schools, etc., and the beat constable maintains around 114 different forms such as servant verification, driver verification, KYC, etc. “All that data is now digitally maintained and has been recorded by beat staff in the last two months in this application,” Patodia explained. The beat cops can also use the app to record details of crimes in the area and identify hotspots for crimes. Data related to time and location of crimes is then used to create a patrolling schedule and take other measures, Patodia said. The e-beat book is also integrated with the 112 helpline so that call dispatchers can push updates about incidents in the area to beat cops.

The question of bandwidth

When we asked Singh how the e-beat book app was able to implement FRT without seemingly running into bandwidth issues, he said, “See, it is implemented but what is the success rate? That search keeps on going, keeps on going, keeps on going.” While fingerprint verification will happen, facial recognition has created problems as it requires large bandwidth, Singh said.

“In both e-beat book and MDT, the facial recognition application is common but it is implementable only once the network is proper. Of course, we have made it in our system that yes, facial recognition will also be implemented but because of this bandwidth problem, both e-beat and this are facing problems so we can’t simply say that we have implemented [the project]. Pilot is different, but actual [implementation] on the field is different. [If on the field, someone asks us,] ‘Please take a snapshot and tell me if a case is registered’, he has to wait for hours together to wait for the result. So that type of commitment, we don’t want to do it,” Singh warned.

Patodia, however, told us that the e-beat book app did not face any bandwidth issues when running FRT. “Almost 5,000 searches [by beat staff] of FRS [facial recognition system] have been done. I think around 8-10 cases have been worked out using FRS and different data searches.” He mentioned that on August 20 itself, beat cops in South Campus were able to nab a person with a stolen bike because he had a dossier and a criminal record.

These 5,000 facial recognition searches are a subset of more than 5 lakh datasets that have been collected via the e-beat book system since the app was launched, Patodia said. A person’s KYC, he clarified, is one dataset in this count. Facial recognition is not used by beat cops to run servant verification or other such routine tasks but to run facial scans of suspected criminals, or potential criminals, or to identify missing children. Assessing the potential for criminality is the “prerogative of the beat officer”, Patodia said.

Senior police officers gave this data to Patodia in their weekly review meetings about the app. 15 people from Mobineers have been deployed within Delhi Police for operational support and maintenance and thus far, they have conducted over 3,000 training sessions for Delhi Police personnel, he said. “As of now, people are quite trained,” he said.

Delhi Police uses facial recognition technology from Innefu Labs

In 2017, a Delhi High Court order directed the Delhi government to procure facial recognition software in coordination with the Delhi Police to track missing children. Delhi-based facial recognition and AI firm, Innefu labs, had won the subsequent Delhi Police tender that sought to buy FRT to identify missing children.

Delhi Police had bought the video and images analytics, one of the five modules of Innefu’s big data analytics product Prophecy, Tarun Wig, Innefu’s co-founder, told us. Innefu Labs is a potential bidders for the National Crime Records Bureau’s Automated Facial Recognition System project.

“I do remember that when we had done the training for them, they had ingested the data of missing children, which was provided to them by the Ministry of Child and Women Welfare and they had matched it with images of children who had been found and put in orphanages. They had got some really good success.” — Tarun Wig, co-founder, Innefu Labs

In a subsequent 2018 affidavit, the Ministry of Women and Child Development had submitted that the software had been used to track about 3,000 missing children. However, in January 2019, the Delhi High Court slammed Delhi Police’s facial recognition system for not cracking any missing children’s case in the previous three years.

Once the technology is handed over to Delhi Police, it is up to them which database(s) they wish to link to Innefu’s technology; Innefu has no control over it. “We are not privy to how they use the system. We don’t have access to the system,” Wig said. After the handover, Innefu’s responsibility is limited to upgrading and maintaining the system. “Beyond that, what they do with the system, we are not privy to it, they don’t allow us access to even the control room unless it is to upgrade or maintain the system which is under their supervision. As a customer, they don’t allow us access to how they use the technology,” Wig said.

Innefu, though, provides regular training to Delhi Police on how to use the system as and when required.

Accuracy of facial recognition system

Wig warned that accuracy rates prescribed in tender documents are for ideal conditions. Accuracy of Innefu’s facial recognition technology, thus, in ideal conditions is above 95% but Wig said that he would not be able to comment in real life scenarios.

Ideal conditions for FRT include front-facing facial scans, adequate environmental conditions such as good lighting, good resolution of the scanning device, etc. When it comes to using FRT on footage from CCTV cameras, it also includes the ideal position of the camera (10-11 feet high), good resolution, and angles that catch the person’s face.

Patodia said that the accuracy of the e-beat book system is currently configured to around 80% so that it shows multiple results. This configuration is under Delhi Police’s control. “They want 90%, they want 95%, accordingly the number of results will vary,” he said. This is because databases have legacy data, that is, data that is at least a few years old, he explained. Increasing the accuracy of the system to 95% or more would preclude a number of results. A fifteen-year-old who went missing five years ago would look quite different today, Patodia said. Echoing Wig, Patodia said that 1:1 matching doesn’t always play out in real-life because of the quality of scans and photos.

Accuracy of the technology itself is a source of concern for Senior Advocate Rebecca John as well. “All across the world, there is so much material and there is so much debate about facial recognition even being a credible science. And therefore, there is a need to first cross those tests before you make it something of a universal application,” John warned. Global debates, she said, suggest that these are not perfect sciences and have high error rates.

John cited the Daubert Standard and Rule 702 of the Federal Rules of Evidence which are used in the US to determine whether the scientific methodology used is valid or not by checking the technique’s reproducibility, error rates, and acceptance within the scientific community. “I think in India, we don’t have those kinds of debates,” John said. Effect on people’s privacy and other issues would come after the question around the science itself is resolved, she said.

Since these technologies are not tested “on the anvil of foundational validity” in India, the scientific community and civil society need to come to an agreement around the technology’s efficacy, she said. “I think this is the time when the scientific community should rise and should bring evidence, well if it is good science, then given evidence of good science, but if it is not, at least get the counterview out. How dangerous these things can be,” she said.

Would the courts allow such facial scans?

The use of facial recognition technology to maintain law and order has not been tested in court as of now, Rahul Mehra, the standing counsel (criminal) for the government of Delhi and Delhi Police, said. “As and when it is tested in the court of law, some jurisprudence will come up,” he said.

Mehra explained that the Delhi Police had started using FRT to trace missing children who may be with NGOs, or in state-run homes, and other places. “Now that was all for a very laudable cause,” Mehra said.

On asking whether people should be able to refuse to get their faces scanned without warrants, John said, “I would imagine so,” but agreed that it would need to be challenged at some stage. “After all, narco analysis was used before the Supreme Court threw it out,” she said.

“With everything, there are rights and there are duties. I have rights, the police also has rights. But those rights are codified rights, those rights are given to them from statutory, from codified statutes,” John said. These rights of the police cannot stem from a Home Ministry notification; “That is not law,” she said. “[T]here should be an act of Parliament which allows you to do that because then we can test the constitutionality of it. It can’t be a random procedure that a police force adopts,” she said.

Such facial scanning “would be in violation of the 9-member judgement of the Supreme Court in Puttaswamy”, John said. Shreya Rastogi, head of litigation and forensics at Project 39A, agreed that Puttaswamy would be “the biggest barrier for the state” but warned that “the legitimate aim of the state is very well made out in case of criminal investigations”.

“It is for the judiciary to decide whether it violated Puttaswamy or any other provision of the law or Constitution. It will have to be tested in the court of law.” — Rahul Mehra, standing counsel for Delhi Police

In fact, general provisions in Indian law give police very wide-ranging powers, Rastogi said. Section 91 of the Criminal Procedure Code (CrPC), she said, allows the court and police officers to “summon any individual to produce a document or thing”. “Now you can interpret the document or thing to be your facial scan,” she said. Similarly, Section 165 of the CrPC empowers the police to carry out searches without warrants, she said.

And when it is tested in court, the argument will hinge around proportionality, which is a “sliding scale”, Rastogi said. “In case of certain investigations, such as riots, we can expect the courts to give police a huge leeway,” she cautioned.

Delhi Police is not the only police force in the country to deploy facial recognition software in the field. Telangana Police, for instance, has been collecting facial and fingerprint data of “potential criminals”. The police use a privately developed app, called TSCOP, to collect facial and fingerprint data from “youngsters” that it thinks might be criminals. All this is based on the police officer’s intuition.

“There’s an immediate need for our courts to become cognisant of this and to actually reimagine these provisions that we have. We have a situation where technology has already moved ahead but the law is yet to catch up. By that I mean we are yet to figure out what are the constitutional and criminal safeguards that we need for the accused as well as the victims alike. What should be there in place and how should this technology be used?” Rastogi asked.

Edited by Aditya Chunduru