<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Davy Uwizera on Medium]]></title>
        <description><![CDATA[Stories by Davy Uwizera on Medium]]></description>
        <link>https://medium.com/@davy_47128?source=rss-44513a83d17b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Up4OWC-w01XPe5Og6QEceQ.png</url>
            <title>Stories by Davy Uwizera on Medium</title>
            <link>https://medium.com/@davy_47128?source=rss-44513a83d17b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:29:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@davy_47128/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Anti-spoof for facial verification in Africa]]></title>
            <link>https://medium.com/smileidentity/anti-spoof-for-facial-verification-in-africa-dd5d28e9a899?source=rss-44513a83d17b------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd5d28e9a899</guid>
            <category><![CDATA[kyc]]></category>
            <category><![CDATA[identity-verification]]></category>
            <category><![CDATA[fraud-detection]]></category>
            <category><![CDATA[anti-spoofing]]></category>
            <dc:creator><![CDATA[Davy Uwizera]]></dc:creator>
            <pubDate>Wed, 04 May 2022 06:52:27 GMT</pubDate>
            <atom:updated>2022-05-04T06:52:27.582Z</atom:updated>
            <content:encoded><![CDATA[<p>The need for “Know your Customer” (KYC) has been on the rise due to increased global demand for digital services. <a href="https://medium.com/smileidentity/beating-facial-recognition-bias-in-africa-ebafb0bf24d7">In one of our previous articles we discussed facial recognition as one approach for KYC and beating facial recognition bias on African faces</a>. In today’s article, we discuss anti-spoof in facial and ID cards verification, to prevent fraud in digital KYC.</p><p>Most companies use digital KYC to trust users with their valuables, which has enabled various companies to provide remote onboarding and services, such as opening bank accounts, registering mobile SIMs, ride services, housing services, etc. But often, in most cases, attackers will try to forge identities to try to bypass KYC technologies. Anti-spoofing is the ability to recognise if the identity presented is real or fake.</p><p>In facial verification, some attacks are simply photos downloaded from social media, but in other cases they could be more sophisticated than that. Below we display some methods used by attackers to spoof facial recognition systems.</p><h3>Selfie Spoof in Facial Verification</h3><p>Selfie spoof, is when attackers are trying to forge the identity of an authentic user using a selfie. There are various ways attackers could do this. Below we display some of these attacks. <em>Faces are partially covered for privacy reasons.</em></p><p><strong>Prints (paper/magazine covers) and soft-copies (screenshots)</strong></p><p>In this attack, the attackers try to present a soft copy of a paper or a screen which has a front-face of an authentic user registered in the system they are trying to compromise. In some cases this type of attack is also tried in the effort of registering a wrong user in the system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*ZWHvQYJVbOsv5-5LpFvzrA.png" /><figcaption>Soft copies and screenshots selfie spoof sample images <strong>(source: SmileIdentity)</strong></figcaption></figure><p><strong>Full and cut-out cardboards</strong></p><p>In some cases attackers may try a big portrait to mimic human body width trying to spoof your system, hence using big cardboards and cardboard cut-outs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*qcWjTamlldBaQ70XwP1IPQ.png" /><figcaption>cardboard spoof sample images <strong>(source: SmileIdentity)</strong></figcaption></figure><p><strong>Deep Fakes</strong></p><p>Recent advances in Artificial Intelligence such as GANS (Generative Adversarial Neural Networks), have made it possible for real time face generating, swapping and transformation. This has resulted in new spoof attacks called deep fakes. Here, the attackers use a program to generate a single or sequence of facial images similar to the authentic user. Below we present a simple Obama deep fake, but the attack could get even more complex and sophisticated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/251/1*D9F-UTUNx1hDCdVQ9m7Tfw.gif" /><figcaption>Deep fake — simple spoof <strong>(source: SmileIdentity)</strong></figcaption></figure><p><strong>Other</strong></p><p>Besides, the discussed types of spoof above which are likely to be often used against facial verification systems, we have seen and dealt with various other types of spoof images including, but not limited to, deliberately blurred &amp; darkened images, cartoons, non-faces, and pets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*WHyE42U5jGT3zwqQvbxxAA.png" /><figcaption>Other types of spoof images <strong>(source: SmileIdentity)</strong></figcaption></figure><h3>ID-cards Spoof in Facial Verification</h3><p>Most of the types of spoof attacks discussed in the previous section such as screen shots, photo-copies, blurred and darkened faces, and deep-fakes apply to ID-cards, see the image below (face and ID-card info are hidden for privacy reasons). The only difference is that for ID-cards, it becomes even much harder to discern since they are usually printed on papers and plastics. And in some cases, it makes them vulnerable to other types of unique attacks such as pasted-faces which we will discuss further in this section.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*kvgxiu3AgR1dJ3rF44homw.png" /><figcaption>Various types of ID-cards spoof images <strong>(source: SmileIdentity)</strong></figcaption></figure><p><strong>Pasted Faces</strong></p><p>When using pasted face spoofs, attackers try to swap an authentic face of an ID-card with a new fake face image. These attacks are often used to illegally create multiple accounts on a system or enrolling with fake information. These kind of attacks could go from just basic pasting of a photo to more sophisticated altered face. Below we show a few examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*77X3tvA7ftt68rtyRV2w-w.png" /><figcaption>Pasted-face spoof images on ID-cards <strong>(source: SmileIdentity)</strong></figcaption></figure><h3>KYC Spoof Attacks Could be Massive</h3><p>If not dealt with the amount of forged identifications and successful attacks on various systems which deal with customer identities could accumulate to big numbers and pose a great threat to user trust on your system.<a href="https://www.reuters.com/article/facebook-misinformation-int/facebook-says-took-down-1-3-billion-fake-accounts-in-oct-dec-idUSKBN2BE12M"> Not long ago, Meta Inc (previous Facebook Inc), mentioned that it deactivated 1.3 billion fake accounts</a> in a period of 3 months. And it seems to be the trend per quarter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-qqIZ5OldScUdr8a8S26RQ.png" /><figcaption><a href="https://www.socialmediatoday.com/news/facebook-publishes-latest-update-on-content-removals-fake-accounts-and-hat-1/605234/">Fake accounts taken down by Facebook each quarter <strong>(source: Socialmediatoday)</strong></a></figcaption></figure><p>Also, <a href="https://www.forbes.com/sites/jeffkauflin/2022/02/02/paypal-admits-45-million-accounts-were-illegitimate-as-fintechs-fraud-problem-grows/">PayPal admitted recently that 45 million accounts were illegitimately using their system</a>. This shows the need and urgency to apply anti-spoof on every KYC process, from the beginning, rather than waiting for cases to accumulate.</p><h3>How to solve the problem?</h3><p>By leveraging recent advances in AI, and using customized best performing deep architectures and loss functions, at Smile Identity we have invested heavily in getting the right data, and using supervised and unsupervised anti-spoof learning to catch all types of spoof discussed and more.</p><p>We use supervised anti-spoof learning to catch known existing types of spoof in our datasets. This means that the machines are taught on a regular basis to recognise by examples various types of spoof attacks and catch them as they try to get into the system.</p><p>For unsupervised learning, we quickly identify potential new types of attacks that are newly tried by attackers against our partners. In this case our machine learning systems are set up to automatically learn new suspicious spoof image patterns on their own and put them in clusters for the human review. Once the suspected new type of attack is confirmed by a human review is then added to known types of attacks, and the process restarts.</p><p>Using our algorithms we have been able to prevent various sophisticated spoof attacks across various partners in Africa. Below we present percent of spoof caught, ensuring our partners safety, per total usage of our end to end KYC system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-W-FKeHC9BupF5j_W7_sRg.png" /><figcaption>Percentage of fraud caught per total number of KYC requests — 2021 <strong>(source: SmileIdentity)</strong></figcaption></figure><h3>Conclusion</h3><p>Facial verification in terms of face similarity matching is not enough for KYC, it should be backed up by a strong anti-spoof system to ensure customer safety. Various types of spoof attacks that are likely to be used against facial verification systems have been presented, using real case spoof attack examples. This highlights how easily unprotected KYC systems could be attacked and bypassed.</p><p>We used facial verification as a use-case, but the same scenarios apply for other types of biometric verifications. Systems without a proper anti-spoof process are likely to accumulate massive fraud which can lead to various criminal activities, take down a company’s reputation and could cost a lot of resources to clean up at any later stage.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd5d28e9a899" width="1" height="1" alt=""><hr><p><a href="https://medium.com/smileidentity/anti-spoof-for-facial-verification-in-africa-dd5d28e9a899">Anti-spoof for facial verification in Africa</a> was originally published in <a href="https://medium.com/smileidentity">SmileIdentity</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beating facial recognition bias in Africa.]]></title>
            <link>https://medium.com/smileidentity/beating-facial-recognition-bias-in-africa-ebafb0bf24d7?source=rss-44513a83d17b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ebafb0bf24d7</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[kyc]]></category>
            <category><![CDATA[facial-recognition]]></category>
            <category><![CDATA[africa]]></category>
            <category><![CDATA[bias]]></category>
            <dc:creator><![CDATA[Davy Uwizera]]></dc:creator>
            <pubDate>Wed, 01 Jul 2020 16:52:16 GMT</pubDate>
            <atom:updated>2020-07-01T16:52:16.786Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beating Facial Recognition Bias in Africa</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/442/1*Mqe9BakIUw_x83-PqdAZrQ.png" /></figure><p><strong>Face Recognition for Digital KYC<br></strong>Global demand for digital services has created a need for digital identity verification. Companies that trust their users (or drivers) with things of value — bank accounts, mobile SIMs, credit cards, motorcycles — often do a “Know Your Customer” (KYC) check to comply with regulation or to prevent fraud.</p><p>In 2020, the Coronavirus pandemic accelerated this, and most KYC checks are moving online as companies try to digitize their day to day operations. The rise of artificial intelligence and big data have enabled innovative approaches to digital KYC, including the use of face recognition (FR). However, face recognition has recently come under scrutiny because of its use by governments and police departments and because of common racial algorithmic bias resulting in lower accuracy for black / high melanin faces.</p><p>The process of developing face recognition involves collecting data and training models on target identities. Results are reflective of the datasets used to train them. But most models available are trained on mostly caucasian faces, so algorithmic accuracy tends to be better on caucasian faces as compared to African and other high melanin faces.</p><p><strong>Bias in the Processing Chain<br></strong>Bias affects the whole image processing chain from phone camera sensors, to face position detection and alignment software, to the recognition algorithms.</p><p>Camera sensor exposure settings are derived from:</p><ul><li>“aperture” (the amount of light allowed to travel through the lens)</li><li>ISO speed (camera sensitivity to light), and</li><li>“shutter speed” (how long light is permitted to hit camera’s sensor)</li></ul><p>Most device manufacturers set them by default to take better pictures on lighter skin. Commercial image standards are also biased in that darker faces are limited to a smaller part of the supported dynamic range.</p><p>And the dataset problem appears again in face position detection and recognition — the geometric facial features learned by face detection models tend to be biased against African and high melanin faces due to the lack of face diversity in widely used training and benchmarking datasets. Commonly used benchmarking datasets like Microsoft’s Celeb-1M and the academic dataset <a href="https://opendatascience.com/the-impact-of-racial-bias-in-facial-recognition-software/">Labeled Faces in the Wild</a> (LFW) are used by researchers to assess model accuracy despite their inherent bias — Celeb-1M has roughly 14.5% African and African-American faces while LFW is estimated to be <a href="https://medium.com/@ODSC/the-impact-of-racial-bias-in-facial-recognition-software-36f37113604c">77.5 percent male and 83.5 percent white</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/885/1*OH1GuxLhSblEsCE3VjEzEA.jpeg" /><figcaption><a href="https://www.standard.co.uk/news/uk/man-stunned-as-passport-photo-check-sees-lips-as-open-mouth-a4241456.html">A black man’s lips mistaken for open by a biased passport checker FR system</a></figcaption></figure><p><strong>Bias in Renowned Face Recognition Providers<br></strong>Given the bias across the chain, it is no surprise that statistics show a huge difference between dark and light face recognition accuracy in some of the leading FR providers. And this bias gets worse when gender is considered. <a href="http://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf">Research done by Gendershades in 2018 on gender classification</a> showed that leading FR providers such as Face ++, Microsoft and IBM, were less accurate on darker faces as compared to lighter ones. The worst results were for images of darker females, where there was a racial accuracy gap between them and lighter male faces of between 21-34% across three major FR vendors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mlys3aDLFBD5JeJ0T68THQ.png" /><figcaption>MIT technology review: Gender-shades face recognition technology evaluation 2018</figcaption></figure><p>While improvements have been made since the original study, especially by Microsoft, a more recent Gendershades review on gender classification in 2019, which included more FR providers, showed that the largest gap in accuracy from major providers was still as high as 31.4%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YM7ADoPAckZuue5E2tucVA.png" /></figure><p>Given that face recognition systems are being used for decisions as critical as law enforcement, these biases can have terrible consequences as the recent <a href="https://www.cnn.com/2020/06/24/tech/aclu-mistaken-facial-recognition/index.html">wrongful arrest of a Black American man due to a face algorithm</a> shows.</p><p><strong>Addressing the Root Problem<br></strong>At Smile Identity, we are working to solve this bias. We make facial recognition services for non-surveillance, consent-based financial inclusion. Our Partners work with us to collect real, front-facing African selfies taken in various fields and activities on a daily basis, making this data unique and representative of day to day life across Africa and for a variety of people with high melanin faces. With this unique data, Smile uses pre-processing techniques customized for African faces and optimized deep transfer learning methods with supervised human review, to improve state of art models’ performance on high melanin faces. Recent experiments conducted at Smile Identity using one of the leading deep learning techniques on our proprietary dataset showed significant improvements in unbiased accuracy. Smile’s algorithm reduced the model gap on dark faces accuracy <em>substantially</em>, with accuracy jumping from 88.0% to 98.9% accompanied by only a modest (0.5%) accuracy loss in mixed LFW faces. We are now working to publish our results in a peer-reviewed study.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ktf7UrjwI0-TtP-m8y3I5Q.png" /><figcaption>Smile Identity “Algo10” Test Results on 100,000 Labeled Faces</figcaption></figure><p><strong>Non-Surveillance Philosophy Embraced by the Tech Giants<br></strong>From access validation to KYC/anti-fraud to anti-trafficking and attendance tracking facial recognition is becoming a common technical component for businesses and governments across the globe. But from the beginning Smile Identity made the decision not to build for law enforcement or surveillance use-cases. Our technology is built for front-facing images and our policies require our Partners to get consent from users before taking their image. <a href="https://www.washingtonpost.com/technology/2020/06/11/microsoft-facial-recognition/">Big companies such as Amazon, IBM and Microsoft have only recently adapted this philosophy</a> while they await Federal U.S. legislation.</p><p><strong>Why We Do This<br></strong>Though solving FR bias is generally a complex problem, by doing this for African and other high melanin faces Smile Identity will unleash a lot of potential in Africa and beyond. And as an African AI engineer working with Smile Identity to eliminate bias in face recognition-based KYC, every day becomes special when you remember that your work is an asset for your continent.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ebafb0bf24d7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/smileidentity/beating-facial-recognition-bias-in-africa-ebafb0bf24d7">Beating facial recognition bias in Africa.</a> was originally published in <a href="https://medium.com/smileidentity">SmileIdentity</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>