While the large part of the new team (6 members now post acquisition) kept focus on innovating on the matching engine, a former team member from Validity and I focused on building the antispoof solution even though this was not a requested feature from the customer. A spoof is the presentation of a counterfeit biometric. With a fingerprint, a spoof can potentially be used to unlock the phone and gain access to the device. Fig 1. above shows sample spoofs we built during the early stages of developing the antispoof solution.
Although the current focus from customers was around ease of use, we knew the focus would quickly shift on security. Instinct and knowing the industry made me feel that this was the right thing to do. Plus, having no antispoof security opened customers to potential security fraud risk which didn’t feel right at all. It didn’t take long to convince the VP of Product and we had a quick prototype which convinced the leadership team as well. In the meantime, as earlier suspected, the prior product of the swipe sensor which had no anti-spoof technology was being tested around the world and criticized for its lack of antispoof solution [8]. This helped convince everyone else on the sidelines.
A two-member team, which included me, led this effort to innovate on the antispoof engine. What followed was a series of inventions that led to a signature product for the company which was a key differentiating factor for the product offering at that time [4][5][6]. Many of these inventions [2][3], were a collaborative team effort with one other inventor and one patent in particular, Systems and methods for a gradient-based metric for spoof detection [1], was a solo contribution and together all the inventions were tied to the Sentry Point Antispoof Engine as a product [4].
Antispoof solution
A spoof is the presentation of a counterfeit biometric, and the quality of fingerprint spoofs can vary greatly while new techniques are being introduced regularly. Because spoofs represent the underlying finger, the matching engine can be fooled into a successful authentication to enable access into the phone. We introduced an antispoof engine into our fingerprint solution as shown above to do a check on whether the presented image is a spoof or an image of a real finger.
With high resolution sensors (>1000dpi), many documented techniques existed in practice which included use of pore detection where fingerprint sweat pores are visible in the images. These pores are less likely to occur with a spoof. With Synaptics sensors, the resolution was 333dpi and with the existing hardware technology of capacity sensing, it was highly improbable to see sweat pores on the image.
But, when doing an exploratory study to determine if sweat pores were visible, I started to notice some anomalies on the raw images that were distinctly captured by our hardware. We started a small capture with spoofs that we constructed in partnership with the testing team and material experts in the team to collect a small sample of prints to analyze further. An intermediate transformation of the image into its gradient form revealed some characteristics that could be potentially exploited to determine the liveness of the image. This was the beginning of the invention.
Invention: Systems and methods for a gradient-based metric for spoof detection
The invention takes an image and computes its gradient image and divides it into blocks.Consider the fingerprint image (left) above with its corresponding gradient image divided into blocks of 8x8 (right). For each block, we compute the histogram of the gradient magnitude and compute the variance in the histogram. Figure below shows the variance in the histogram plotted as an image where a dark block represents high variance and a lighter block shows low variance in the block.
What I found was that for many of the spoofs, depending on the material of the spoof, the variance was spread out. However, this was less so with the images from real fingers. When we plotted the histogram of these variances, a clear distinction stood out. Below is an example of the histogram of variance for a real finger and a spoof for a large fingerprint image where the feature is much more distinct.
Figure below shows the variance in the histogram plotted for a full fingerprint image for a real finger and a spoof, where a dark block represents high variance and a lighter block shows low variance in the block. The real finger shows higher variance in the histogram vs the spoof image.
The variance of these histograms or some distance measure of the histogram like the Earth Mover's Distance (EMD) became a strong feature in the final antispoof engine. The final solution on the antispoof engine took several hundreds of features we computed from various intermediate representations on the image and fed into a neural network to compute a final result for liveness. The gradient feature described above [1], was one important and unique feature for which I was the sole inventor.
Conclusion
The full solution was by no means a small endeavor and it involved multiple teams from hardware, software, material experts and QA teams to try and break the antispoof engine. Geographically, we had teams in India, China and San Jose all in collaboration with us on this effort. As in any production solution, this was a team effort for the full solution and I was fortunate enough to innovate in this space.
The antispoof engine was the first of its kind on the mobile phone and was showcased at the Mobile World Congress in 2016. Competitors like Precise biometrics found alternate solutions in Feb 2017, a year later giving Synaptics a lead time of more than a year in the market with this offering [7]. This helped Synaptics offer a solution to the market for spoofs and ensure security on the device against spoofs and helped win the business with multiple tier 2 customers beyond Samsung like Alcatel, ASUS, Xaomi and many many others. As of 2016, 200 million fingerprint sensors were shipped with the full solution.
References
Rohini Krishnapura, ‘Systems and methods for a gradient-based metric for spoof detection’, Issued Jan 8, 2019, US Patent number: 10176362, Assignee: Synaptics Incorporated
Rohini Krishnapura, Anthony P. Russo, ‘Systems and methods for spoof detection based on local interest point locations’, Issued November 13, 2018, US Patent number: 10127429, Assignee: Synaptics Incorporated
Rohini Krishnapura, Anthony P. Russo, ‘Systems and methods for improving spoof detection based on matcher alignment information’, Issued November 6, 2018, US Patent number: 10121054, Assignee: Synaptics Incorporated
Synaptics Adds Proprietary Anti-Spoofing to SentryPoint Security Suite | Synaptics Incorporated, Feb 2016
Area Touch and Swipe Fingerprint Sensors | Natural ID, Synaptics
Protecting Against Fingerprint Spoofing in Mobile Devices, Synaptics White Paper, 2016
Precise Biometrics software in Xiaomi Redmi Note 4X 16 February 2017 13:40 GMT
Samsung Galaxy S5 fingerprint scanner already hacked using 'faux fingerprint'
