Shaping public safety: IDEMIA’s take on biometrics

Image created by DALL·E 3.

Public safety is a shared responsibility, with governments, enterprises, and individuals collaboratively working towards the common good. To this end, technology plays a pivotal role by improving efficiencies, optimising resources, and driving better results more quickly.

From road safety, to law enforcement aid, IDEMIA seeks to enhance public security with its smart solutions, central to which is the use of biometrics.

Frontier Enterprise had the opportunity to sit down with Vincent Bouatou, Deputy CTO, Head of Strategic Innovation & IP at IDEMIA Public Security, for an exclusive interview about the integral role of biometrics in the company’s public security strategy.

- Advertisement -

What’s currently happening over at IDEMIA’s R&D labs?

At IDEMIA, I serve as the Deputy CTO for the Public Security Division, which encompasses four business lines: law enforcement and public security, travel and transport, road safety, and smart biometrics. AI is fundamental across all these areas, forming the basis of our value proposition. Essentially, we offer solutions that automate key processes to enhance efficiency and security.

One notable value proposition is our ability to automate law enforcement processes. We are very excited about the recent developments in enhanced latent fingerprint testing, or ELFT. This involves an automated system’s capability to analyse a latent fingerprint — often partial, smudged, and of low quality — from a crime scene and match it against a large database of criminal records. We emerged as the clear leader in this benchmark, significantly ahead of the competition.

The system is evaluated on two critical metrics: its accuracy in correctly identifying the right person in the criminal database, and its ability to flag which searches warrant further investigation. In criminal identification, while the final decision rests with a human latent fingerprint examiner, the system’s role is crucial. It not only identifies potential matches, particularly useful for first-time offenders not previously in the database, but also advises when a match is unlikely, thereby saving examiners from pursuing fruitless leads.

Vincent Bouatou, Deputy CTO, Head of Strategic Innovation & IP, IDEMIA Public Security. Image courtesy of IDEMIA Public Security.

Given that latent fingerprint examiners spend approximately 80% of their time reviewing negative results, the ability to efficiently flag only the relevant searches is a significant time-saver. This efficiency is crucial, as training a latent examiner takes months, if not years. So being able to optimise that workload is very important to them.

Another innovation we’re showcasing is a free-flow identification system for air travel. Traditionally, passengers proceed one-by-one through boarding checkpoints where a system or a person verifies that the names on their tickets and ID documents match, allowing them to board the plane. Our new system facilitates a continuous flow of passengers through checkpoints without stopping. It uses facial recognition technology to check each passenger’s identity as they walk through the gate.

So it’s not a queue, it’s not one by one? It’s all at once?

Yes. We believe this has the potential to really change certain operations, perhaps even border control operations in the more distant future, as recently highlighted by the ICA Commissioner. There’s still considerable work to be done before border control authorities are comfortable with people crossing borders in this manner. However, for boarding operations, which are somewhat simpler, we envision a near future where this could be implemented.

What was the bottleneck in achieving this in the past? Was it the efficiency of the algorithm, was it the processing speed, or was it an operational problem?

In the past, the main challenge with boarding was the quality of technology and algorithms. Recently, advancements have enabled us to detect, track, and identify multiple individuals in real time with just one camera. This capability was not available two or three years ago. It’s really about the ability to process multiple identities on the fly that’s really changed things.

A significant breakthrough in this area was the development of the YOLO algorithm — “You Only Look Once.” It allows us to detect all faces in a single image or video stream simultaneously, without needing to analyse every single pixel individually. Instead, it takes a more holistic approach to image analysis, which has been a pivotal moment, enabling us to consider mass processing in parallel.

What about private CCTV cameras that are not under government control? What kind of developments have we seen in that arena, especially with poor quality cameras? Are we still able to get better data out of them?

In the context of criminal investigations, if an incident occurs and there are cameras present, we often have to work with footage from cameras that might be 10-15 years old. These cameras, once considered high-quality, now produce relatively low-quality video. Despite this, we’ve made significant advancements in extracting valuable information for law enforcement agencies from such low-resolution footage.

Typically, investigators will input available video into our system, which then indexes and interprets the content, providing insights into what occurred and identifying involved individuals. Our system can link faces detected in extensive video footage —sometimes up to 100 hours — to criminal databases, identifying known individuals at the scene, their interactions, and their movements over time, all integrated into a timeline. This technology greatly aids investigators by streamlining the analysis of abundant video data, which is crucial during investigations.

In cases of major incidents like terrorist attacks, law enforcement might need to sift through thousands of hours of footage. Traditionally, this would involve setting up a large-scale operation, perhaps in a warehouse, with up to 200 operators tasked with processing all that data. Even then, coordinating communication and information sharing among so many investigators poses a challenge. However, with the technology that we’ve developed, you’re able to get all of the non-important stuff filtered out by the system and have a handful of investigators really focus on what’s important within that data.

In road safety, we are leveraging AI to detect new types of violations, enhancing what has traditionally been focused on speeding. Historically, our AI was primarily used to identify licence plate numbers. However, we have now expanded our capabilities to detect specific actions such as unauthorised U-turns, and even activities within the car, like whether someone is wearing a seatbelt or using their phone while driving. We also use AI to check for compliance in carpool lanes, ensuring that there are multiple people in the car if it is using a carpool lane. This broadened application of AI within our road safety solutions underscores its growing importance as the foundation of our value proposition.

Do you see biometrics as sort of the linchpin, or the single most important factor in identity in the future?

The short answer is yes. Biometrics is currently the only method that can automatically, demonstratively, and safely link a government-issued credential and identity to an individual. We’ve been doing this manually for centuries.

Biometrics has been used for millennia, manually. We’ve simply automated that process. We’ve reached a point where, due to the ever-increasing volumes of travellers and crime investigations, automating this process is absolutely necessary. While full automation is possible for lower-sensitivity cases, sensitive scenarios such as criminal investigations will likely always require a human to confirm hits or matches. Nevertheless, biometrics is the cornerstone in automating all these applications so that we can cope with these large volumes.

That being said, there’s still a lot that we are doing to help even if we keep manual adjudication as the final decision. We maintain strict control over our false positive identification rate and regularly have it evaluated by third parties to ensure its accuracy. For example, we confirm that a reported 0.1% false positive rate genuinely means 0.1%.

Additionally, we focus intently on demographic bias to ensure that our criminal identification practices do not unfairly target any demographic group, thereby increasing the likelihood of equitable outcomes. We regularly submit our algorithms to NIST (National Institute of Standards and Technology), which evaluates them for demographic effects. NIST has confirmed year after year that IDEMIA has effectively minimised demographic impacts, maintaining a very low influence of demographic groups on our performance. This is something I am particularly proud of, as we have been active in nearly 180 countries for almost 50 years, long aware that demographic effects should not influence our solutions. It has taken a long time for the community to recognise the importance of integrating these considerations into metrics, evaluations, standards, and requirements. But now, customers are beginning to include fairness requirements when procuring biometric systems, which marks a huge step forward.

We have been actively involved in developing these standards, alongside other industry stakeholders, and have also ensured that the recently issued European Artificial Intelligence Act includes provisions for high-risk use cases to demonstrate bias mitigation. This is a huge win for everyone, and we’re very happy because we’ve been working on these issues for years and believe we are well-positioned to guarantee this level of performance to our customers.

With AI integral to your business, how do you process your customer data from all over the world?

We are active in numerous countries and work with high-profile clients, which allows us to access analytics that monitor system performance. These analytics ensure that our systems perform according to specifications and do not degrade over time. However, training presents a more delicate situation. We never directly access our customers’ operational data. In specific instances, we may have the system operator, typically not us, run diagnostics and report back, but direct access to the data is avoided.

We usually train with data that we have collected ourselves. One of the reasons we have this data is because we’re also developing and manufacturing sensors, and when you want to design, develop, and fine-tune a sensor, having people walking towards the sensor is necessary.

We’ve also been developing synthetic data generation algorithms. While these cannot be used for biometric training, they are invaluable for testing biometric systems in production. Working with government agencies, who are particularly cautious about ensuring operational data is not compromised, limits our ability to directly access the systems.

So, we provide large synthetic datasets that enable our customers to conduct endurance and throughput tests on systems to ensure they meet specifications. However, synthetic data is not used for training purposes, except in road safety. For example, since a car’s appearance is consistent, we can use synthetic imagery to train algorithms to detect cars and measure their speed. However, for humans, with the current state of technology, we don’t find synthetic data very useful for training.

You can’t imagine how much testing we’ve done on synthetic biometric data. Even though we use it primarily for system testing, we must ensure that no personal information from the testing data leaks into the synthetic data. We have extensively tested to confirm that no personally identifiable information from the natural data we use for training contaminates our synthetic data. Additionally, we rigorously verify that the synthetic data we generate accurately represents the data we aim to emulate.

We made numerous tests to ensure accuracy, but we also want to make sure we are actually creating synthetic people in the sense that even if parts of my face were used in the training of our generative AI algorithms, no identifiable part of my identity is evident — even in a very diffused manner — in the synthetic data we generate.