How Apple can improve Face ID’s reliability – Info Gadgets

After some serious instances that raised privacy concerns, encryption and security have become top priority for the technology giants. After the introduction of biometric authentication in mobile phones, it has become a huge challenge for phone makers to create a biometric method that meets the following criteria:

• Ease of use
• High reliability
• Support enhancing the design and aesthetics of the phone

Apple has come up with a solution that would be a union of all the three factors and, they call it the ‘Face ID’. As the name suggests, Face ID uses our face as biometric-login to unlock iPhones. Unlike earlier facial recognition systems, Face ID uses depth information coupled with advanced machine learning algorithms, making it more trustworthy.

Despite using this advanced technology, the Face ID system is not foolproof. No prediction system can be 100% accurate, but since this is a matter of privacy and access to sensitive information including payments, we need a solution with the best possible reliability. In this article, I’ll mention some failed cases followed by a briefing on how Face ID works. Finally, I provide some suggestions that Apple can incorporate to improve Face ID’s prediction.

The following are tested scenarios where the system falsely predicts:

• The Face ID fails to distinguish between me and my father[1], that motivated me to write this article
• The system also struggles to distinguish among some triplets[2]

How Face ID works [3]:

Face ID works by the integration of three major hardware components: an IR projector, an IR camera and a flood illuminator. The flood illuminator produces electromagnetic waves in the Infrared(IR) spectrum that is projected onto our faces by the projector. The IR radiations are invisible to our naked eye and are widely used in imaging during nights and in some coveted operations, enabling the Face ID system to work perfectly in dark as well. These electromagnetic radiations bounce and reflect off our face and are captured by the IR camera. The built in software then reconstructs the data acquired by the camera and reproduces a 3D model of our face that is fed to the Neural Engine Chip. This chip makes powerful computations quickly to make predictions of our Face.

Key advantages of Face ID :

• Invariant to various lighting conditions, including darkness
• Invariant to our eyewear
• Invariant to different orientations and depths that we hold our phone with
• Adapts to our natural ageing process

Hardware improvements:

As mentioned earlier, the Face ID system relies heavily on the data acquired by the hardware components. Although the current resolution of the dot projector is high (~30,000 dots), it is likely that it may not be high enough for the system to distinguish similar faces that have subtle differences. Hence Apple should work to increase the resolution of both the dot projector and of the IR camera.

Software improvements:

Hardware components aside, higher accuracy can be achieved even by enhancing just the relevant software package. For instance, depth effect in Google’s Pixel phone was achieved only through software, as against, using dual camera hardware in iPhone X and Samsung’s flagship phones. In addition, it is relatively easier to implement software updates for existing and future devices rather than hardware upgrades that can only be applied only to future products.

Over the past couple of years, Apple has published a few research papers about the technologies that power Apple devices like the ‘Hey Siri’ feature. However, Apple has not published a research paper on how Face ID system makes its prediction. Nevertheless, since they advertised that it uses Machine Learning and a dedicated neural engine chip to make these predictions I reckon they use a deep neural network with 3D convolutions. Also, the advantages of the Face ID system discussed above, suggest that, the algorithm is based on deep learning.

Having this in mind, the following are a few suggestions that Apple can incorporate to bolster their Machine Learning Prediction Model:

1) Data augmentation

One big issue with Face ID is misclassification of family members. So Apple should collect a lot of data of twins and parents with their children. This is because, the performance of the system, that uses deep learning model, increases with increase in the size of the training dataset.

2) Inspect the Machine learning model

After data acquisition, they should focus on increasing performance on the augmented dataset. The first thing Apple can do is visualize different feature maps,[4] while feeding the true- negative/false-positive cases.

3) Tinker the model

With the augmented dataset, and given the reason of failure to authenticate the right face, Apple can focus on changing model’s hyper-parameters to boost the training and test accuracy. One good idea is to go deeper into higher dimensions. This might be particularly helpful in classifying similar classes like twins. A simple and clear illustration by Colah[5] suggested that two indiscernible data points can be separated in higher dimensions using non linear functions. A similar strategy can be applied in deep learning, i.e. increasing the number of layers (going deeper) and/or by trying different nonlinear functions.

4) High threshold

Since face detection is a scenario of matching the owner’s image with the data already fed, they are likely to use a threshold value to verify the owner’s face. One thing that Apple should realize is, it is ok if our face is not recognised in certain situations; however, it is risky if it unlocks the phone for someone else’s face. Therefore Apple should increase the threshold value.

Like touch ID that is ubiquitous in Apple ecosystem, Face ID is likely to make its way to iPads and Macbooks. I hope Apple can work on this before it introduces the next generation Apple products and iOS. After all who doesn’t want a gadget that is not only cool but reliable too?

References:

1) How Apple can Fix FACE ID link

2)iPhone X Review: Testing (and Tricking) FaceID by Wall Street Journal link

3) How Face ID Works On iPhone X — Forbes Link

4)Visualising and Understanding Convolutional Networks, Matthew D Zeiler, Rob Fergus link

5) Neural Networks, Manifolds, and Topology link