- Email: [email protected]
Subaru showcases how biometrics can cut car deaths
NEWS provide the foundation for collaboration around future technical specifications. As well as formulating a vision for web payment security, its charter includes developing use cases, providing gap analysis, and identifying new standardisation opportunities for each of the three organisations. The technical work involved will be carried out by member organisations. For more information about joining the Web Payment Security Interest Group, visit http:// www.w3.org/securepay/. The FIDO Alliance’s mission is to change the nature of authentication with open standards that are more secure than passwords. In September 2018, it launched a Biometric Component Certification Programme which verifies that biometric systems meet recognised standards for recognition performance and presentation attack detection (PAD) and are fit for commercial use.
automotive market
Subaru showcases how biometrics can cut car deaths
J
apanese car maker Subaru has been showcasing and winning plaudits for new vehicle models that use facial recognition technology to check the driver is not tired or distracted – a major cause of traffic accidents and deaths. The latest Subaru 2020 Outback SUV, launched last month at the New York Auto Show, includes a new DriverFocus ‘distraction mitigation’ system which uses a dedicated infrared camera and facial recognition technology to identify signs of driver fatigue or distraction, and provide audio and visual warnings to alert the driver and passengers. According to US Government figures, nearly 3,500 people in the US were killed in crashes involving distracted drivers in 2016. Worldwide, close to 1.25 million people are killed in road crashes annually, with the US Government estimating that over 90% of crashes are caused by human error – including distraction and fatigue, driving too fast, misjudging other drivers’ behaviour and alcohol. Last month, Subaru’s Ascent and Forester car models were also named among the ‘10 Best Family Cars of 2019’ by Parents Magazine, based on their safety, convenience and value. The Forester includes the DriverFocus face recognition system, and both models offer AI safety systems such as automatic pre-collision braking, lane departure and sway warning technology. The attention and awards given to Subaru highlight the role that biometric systems
May 2019
Subaru 2020 Outback: dedicated facial recognition technology identifies driver fatigue or distraction.
increasingly play in the development of ‘smart’ and eventually driverless vehicles. Parents editor-in-chief Liz Vaccariello emphasised the importance of AI-based safety in its top 10 choice. “In developing this list, we wanted to evaluate what’s important to families at a level that exceeds other car rankings,” she said. “This meant understanding the fundamental features needed in a vehicle, while keeping safety top of mind.” Subaru US president and chief operating officer Thomas J Doll said: “The Parents 2019 accolade underscores our commitment to providing customers with vehicles made to the highest standards of safety.”
racial bias
US study: better image quality could cut face system bias
A
research team from the Florida Institute of Technology and the University of Notre Dame have shed new light on the controversy surrounding the racial bias detected in face recognition technology. In their paper ‘Characterizing the Variability in Face Recognition Accuracy Relative to Race’, published last month, the researchers explore why inequalities in face recognition accuracy occur, and what could be done to mitigate them. Their findings suggest that facial recognition systems may be improving, but that quality of images is a major issue. The five researchers – Krishnapriya K S, Kushal Vangara, Michael C King, Vitor Albiero and Kevin Bowyer – tested the face recognition accuracy between African-American and Caucasian image cohorts of the MORPH dataset using four face matchers – two commercial matchers, COTS-A and COTS-B, and two CNN (convolutional neural network) systems, VGG and ResNet. They report that, for all four matchers, the African-American image cohort had a higher false match rate and lower false non-match rate. They also found that ROC curves compare
verification rates at the same false match rate, but that the different cohorts achieve the same false match rate at different thresholds. “This means that ROC comparisons are not relevant to operational scenarios that use a fixed decision threshold,” they report. The researchers add: “Using ICAO compliance as a standard of image quality, we find that the initial image cohorts have unequal rates of good-quality images.” In their conclusion, the authors suggest that image quality is a key issue, stating: “We find no good evidence for a difference in the face detection or failure-to-enrol rate between the AfricanAmerican and Caucasian cohorts. We find the African-American image cohort is disadvantaged on FMR and advantaged on FNMR compared to the Caucasian image cohort. Across a set of two COTS matchers and two well-known CNN matchers, two matchers have a better ROC for the Caucasian cohort and two have a better ROC for the African-American cohort.” They add: “The more important point is that ROC curves are not an appropriate way to compare face recognition accuracy across demographic cohorts. When ICAO compliance is used to select subsets of the images that are more equal on image quality, we find that the low-similarity tail of the genuine distribution is reduced for both cohorts.” Tackling the question of why their results are at odds with those of previous works, they state: “Our results obtained using the COTS-A and VGG matchers broadly agree with previous works. Our results obtained with the COTS-B and ResNet matchers show that the pattern of ROC results seen in previous works is not general across all matchers. “At least two factors may be involved. One is that COTS-B and ResNet are newer matchers than those used in previous works, and face recognition technology has improved over time. A second factor is that previous works did not identify accuracy differences at the level of the impostor and genuine distributions. If previous works had identified the pattern of the AfricanAmerican cohort having higher FMR combined with lower FNMR, it may have suggested that the ROC for the African-American cohort could in principle be better or worse. More fundamentally, ROC curve comparisons are not an appropriate way to compare accuracy across demographic cohorts for an operational scenario that uses a fixed decision threshold.” They conclude: “Examination of the distributions of scores for elements of the ICAO compliance check suggests that the distribution of image brightness scores is a big, perhaps the main, factor. At this time, we can only speculate on why the African-American image cohort has a larger fraction of poorly-lit images. One speculation is Continued on page 12...
Biometric Technology Today
11
automotive market
Subaru showcases how biometrics can cut car deaths
J
apanese car maker Subaru has been showcasing and winning plaudits for new vehicle models that use facial recognition technology to check the driver is not tired or distracted – a major cause of traffic accidents and deaths. The latest Subaru 2020 Outback SUV, launched last month at the New York Auto Show, includes a new DriverFocus ‘distraction mitigation’ system which uses a dedicated infrared camera and facial recognition technology to identify signs of driver fatigue or distraction, and provide audio and visual warnings to alert the driver and passengers. According to US Government figures, nearly 3,500 people in the US were killed in crashes involving distracted drivers in 2016. Worldwide, close to 1.25 million people are killed in road crashes annually, with the US Government estimating that over 90% of crashes are caused by human error – including distraction and fatigue, driving too fast, misjudging other drivers’ behaviour and alcohol. Last month, Subaru’s Ascent and Forester car models were also named among the ‘10 Best Family Cars of 2019’ by Parents Magazine, based on their safety, convenience and value. The Forester includes the DriverFocus face recognition system, and both models offer AI safety systems such as automatic pre-collision braking, lane departure and sway warning technology. The attention and awards given to Subaru highlight the role that biometric systems
May 2019
Subaru 2020 Outback: dedicated facial recognition technology identifies driver fatigue or distraction.
increasingly play in the development of ‘smart’ and eventually driverless vehicles. Parents editor-in-chief Liz Vaccariello emphasised the importance of AI-based safety in its top 10 choice. “In developing this list, we wanted to evaluate what’s important to families at a level that exceeds other car rankings,” she said. “This meant understanding the fundamental features needed in a vehicle, while keeping safety top of mind.” Subaru US president and chief operating officer Thomas J Doll said: “The Parents 2019 accolade underscores our commitment to providing customers with vehicles made to the highest standards of safety.”
racial bias
US study: better image quality could cut face system bias
A
research team from the Florida Institute of Technology and the University of Notre Dame have shed new light on the controversy surrounding the racial bias detected in face recognition technology. In their paper ‘Characterizing the Variability in Face Recognition Accuracy Relative to Race’, published last month, the researchers explore why inequalities in face recognition accuracy occur, and what could be done to mitigate them. Their findings suggest that facial recognition systems may be improving, but that quality of images is a major issue. The five researchers – Krishnapriya K S, Kushal Vangara, Michael C King, Vitor Albiero and Kevin Bowyer – tested the face recognition accuracy between African-American and Caucasian image cohorts of the MORPH dataset using four face matchers – two commercial matchers, COTS-A and COTS-B, and two CNN (convolutional neural network) systems, VGG and ResNet. They report that, for all four matchers, the African-American image cohort had a higher false match rate and lower false non-match rate. They also found that ROC curves compare
verification rates at the same false match rate, but that the different cohorts achieve the same false match rate at different thresholds. “This means that ROC comparisons are not relevant to operational scenarios that use a fixed decision threshold,” they report. The researchers add: “Using ICAO compliance as a standard of image quality, we find that the initial image cohorts have unequal rates of good-quality images.” In their conclusion, the authors suggest that image quality is a key issue, stating: “We find no good evidence for a difference in the face detection or failure-to-enrol rate between the AfricanAmerican and Caucasian cohorts. We find the African-American image cohort is disadvantaged on FMR and advantaged on FNMR compared to the Caucasian image cohort. Across a set of two COTS matchers and two well-known CNN matchers, two matchers have a better ROC for the Caucasian cohort and two have a better ROC for the African-American cohort.” They add: “The more important point is that ROC curves are not an appropriate way to compare face recognition accuracy across demographic cohorts. When ICAO compliance is used to select subsets of the images that are more equal on image quality, we find that the low-similarity tail of the genuine distribution is reduced for both cohorts.” Tackling the question of why their results are at odds with those of previous works, they state: “Our results obtained using the COTS-A and VGG matchers broadly agree with previous works. Our results obtained with the COTS-B and ResNet matchers show that the pattern of ROC results seen in previous works is not general across all matchers. “At least two factors may be involved. One is that COTS-B and ResNet are newer matchers than those used in previous works, and face recognition technology has improved over time. A second factor is that previous works did not identify accuracy differences at the level of the impostor and genuine distributions. If previous works had identified the pattern of the AfricanAmerican cohort having higher FMR combined with lower FNMR, it may have suggested that the ROC for the African-American cohort could in principle be better or worse. More fundamentally, ROC curve comparisons are not an appropriate way to compare accuracy across demographic cohorts for an operational scenario that uses a fixed decision threshold.” They conclude: “Examination of the distributions of scores for elements of the ICAO compliance check suggests that the distribution of image brightness scores is a big, perhaps the main, factor. At this time, we can only speculate on why the African-American image cohort has a larger fraction of poorly-lit images. One speculation is Continued on page 12...
Biometric Technology Today
11