Why you should use multimodal biometric verification for security systems
Deep Neural Networks, TensorFlow, OpenCV, Librosa
Applying machine learning techniques to biometric security solutions is one of the current artificial intelligence trends. This article will cover multimodal biometric recognition system development, and elaborate on why unimodal biometric identification isn’t a reliable model.
At a core level, the majority of biometric data can be classified into one of two categories:
There’s behavioral biometrics, which comprises gestures, handwritten text, vocal recognition and walking patterns.
And there’s physical biometrics, which comprises fingerprint, vein, iris and facial recognition, as well as the geometry of the hand.
Data science consultants can use ML’s ability to mine, search and analyze big sets of data for improving biometric security system performance and reliability.
Face Recognition
The facial recognition algorithm analyzes 64 reference points on the human face, such as the horizontal and vertical length of the nose, the width of the nostrils, the dimensions of the chin, the location of the cheekbones and the forehead, and the horizontal distance between eyes.
Next, a dimensional descriptor vector of facial features is formed and compared to database images by determining the Euclidean distance between this new face and each other face in the database. If a match exists, it will be flagged.
A problem exists, though. In the age of social media and other open sources of photographs, it can be trivial to obtain a picture of a person. One of the major security factors in biometric recognition is confirming that a user’s face is ‘alive’.
There are two primary strategies needed in facial recognition:
- First, liveness detection — determining that the picture in question is not any type of mask or an altered image, and using anti-spoofing tech to confirm that the person before the camera is a live individual.
- Second, comparing the picture with a formal identification — cross-matching the selfie picture submitted against some kind of formal ID. This method uses optical character recognition technology to check text on the picture against an official form of identification. The data that can be accessed on a formal identification document can serve as a valuable indicator.
Voice Recognition
Among biometric verification systems, voice verification is one of the most cost effective, as it only needs an application capable of speech classification and a microphone to capture the speech.
In order to produce a high quality voice sample for verification, the user should be asked to state some passphrase consisting of either words or numbers. Algorithms that process voice recognition are able to discern similarities and differences among characteristics and patterns in the voices of individual speakers. Once a voice sample is read in, it can be directly compared with voice models which have previously been parsed and stored within the pre-existing database.
It’s important to note that the current state of voice recognition technology isn’t 100% accurate or perfect. A person’s voice alters due to factors including their age, their health, and even their mood over time.
But artificial intelligence in the field of biometrics is able to assist with this. It’s possible to train a system by working with commonly occurring distortions in speech, as well as routine background noise. A neural network is able to learn from encountering those types of spurious inputs and filter them out when analyzing a voice sample.
Optical Character Recognition
OCR, a classic machine learning and computer vision task used to distinguish printed or handwritten text characters inside digital images of physical documents, is not an official biometric verification technology. But it can be used as a preliminary tool for comparing the provided information with user documents, for example, for drivers licenses recognition.
Developing Multimodal Identification System
In the table below you can find the results of my research on biometric recognition solutions comparison.
It proves that none unimodal biometric solution is capable of providing the level of security high enough to avoid hacking. That’s why a multimodal solutions should be used.
A system featuring multimodal biometric recognition requires the confirmation of at least two identifiers, which results in a higher level of accuracy.
Let’s overview a system based on voice and face recognition technology, and is accessed by a mobile app.
At the first stage, a potential user will create a photographic imprint, which will be stored via camera on the device in question. The OpenCV library will assist in converting and normalizing the biometric imprint.
The system will then parse the photo, using the 64 facial landmarks. Once mapped, the landmarks and an overall cutout image of the user’s face will be fed into a deep neural network. This network has been taught with the aid of the TensorFlow library.
Once the neural network has processed the facial image, it will create an eDNA feature vector, which represents all the major biometric characteristics of the user’s facial image. An eDNA vector generally measures approximately 2048 bits.
Now, the eDNA vector can be compared with the record already on file within the system database. This method can’t be reverse-engineered, as any potential hacker has no ability to access that original record in the database.
Finally, this system will take an updated record of the user’s image at intervals, keeping up with the user’s potentially changing appearance.
At the second stage the user gives a vocal sample via microphone. The Librosa library takes in the audio sample, converts it into a workable form and then sends biometrics to the neural network (DNN).
Once again, a 2048 eDNA vector will be created, factoring characteristics like intonation, tempo, timber, pitch and other aspects on which a neural network has been taught.
The idea of verifying user identity through a multimodal combination of biometric data is increasing in popularity. Driving this upswing in adoption is the need for ever-better accuracy in recognizing users while defeating spoofing or hacking attempts.
Biometric verification driven by machine learning technology will be one of the key trends which determine the future of artificial intelligence for 2020 and beyond.
