General state of AI systems in 2022
The accuracy of AI systems has vastly improved since 2020. However, many AI systems still use outdated libraries and services. And some of these are open source. Such outdated libraries and services can lead to a dip in accuracy. are several options for face recognition and OCR, but it can compromise heavily on accuracy, depending on how robust and well maintained the libraries are.
Why accuracy matters
In any industry such as FinTech, Crypto, gaming etc., onboarding is a crucial step that you cannot do without; every user must go through it. And this includes several aspects of face recognition here such as Liveness check and face matching that cannot be bypassed. And because every user must go through such verification, it can even affect user retention depending on how user-friendly it is. AI/ML code helps arrive at an accurate decision faster. As far as a business is concerned,
AI and its level of accuracy is the difference in onboarding a few hundred or many thousands in the same period. For a smaller business handling a couple of requests a minute, it may not make a difference but as you scale and handle hundreds of requests a minute at times, the speed to accuracy becomes a critical factor. Another aspect of low accuracy face recognition systems that use AI is randomly accepting fraudsters or rejecting genuine customers. This is why high accuracy AI engines matter; to let the right ones in, and the bad ones at bay.
Recent trends of inaccurate face recognition in law and order
AI systems are not free of racial bias. This is a problem that hinders the usage of face recognition even by government institutions, who actively need some form of face recognition. The degree of racial bias in face recognition depends on the AI and the face recognition engine itself, apart from other factors such as lighting, internet connectivity, and image processing. Speaking of racial bias and its implications, let us consider the case of Microsoft’s face recognition software Face API. Unions in the UK claimed that it does not accurately identify black drivers on the Uber app. Following misidentification, several black drivers even had their accounts deactivated. Because of misidentification by inaccurate AI, companies and governments are reinspecting whether face recognition must be used.
Risks of OCR for identity verification
There is less debate on whether OCR must be used. Many organizations, particularly real estate, legal firms and FinTech are now benefiting greatly from it. But to have great OCR, the AI model must be very well trained. Now let us look at how OCR is hampered by poor AI. OCR without AI or backed by AI with low accuracy may not be able to determine the structure of a document. OCR may not work properly, and the risk of duplicates in the system (if cross verification is not done) goes up. Image rectification may need to be done, which requires good machine learning capabilities before OCR can happen. You may be tempted to try generic OCR engines such as Google/Amazon. However, they are not trained for specific documents, and have lower accuracy. They work great for generic use cases, not for specific use cases.
Rise in cost due to low accuracy of AI (FR+OCR)
There are several reasons why there is a rise in cost due to low accuracy of AI. Some of them are mentioned below:
- During the process of onboarding, due to increased TAT, customers may just drop off. When this happens, and the customer acquisition cost is high, we lose out on the CLTV of a customer, and at scale this hits a business really hard.
- In a face recognition system, the implicit racial bias and resulting inaccuracies can lead to lawsuits that go up in the millions. If you decide to review spurious cases, then manual reviews will add to the cost. This will affect user experience also.
- In an OCR engine with low-accuracy AI, there will be duplicates and inaccurate entries both of which lead to the need for heavy manual verification and increase in time and cost.
- If the AI is slow, it can further cause the FR or OCR system to slow down, reducing the number of applications or claims processed, causing drop offs.
- If the AI system is used for fraud detection, and is not accurate enough, then this will prove an easy access point for hackers, leading to losses in the millions of dollars for companies.
Build vs. Buy: The dilemma
Many organizations, big and small, go through this dilemma of build vs. buy. They start with a thought if it is possible to build a face recognition system in-house. But what they underestimate at times is the difficulty in finding good AI engineers. And even if they do, developing those models is a tough task. Not to mention, finding valid sample data for processing and later updating AI models to combat evolving frauds and to support different types of documents. You will need a separate product team to handle all the resources related to it. And what happens when you buy? Companies can focus on their core business, and let an expert take care of the AI models and face recognition. This is the case, unless building computer vision and capable image processing (using AI technology) serves as a strategic differentiator to their core business. Having an organization with the right expertise partnering will help drive more ROI as you scale.
How to get around these problems?
Now that you have decided to buy. How do you make sure that you are going in for the right service provider for your face recognition feature? These are some of the things you can do:
- Opt for an AI-based identification services provider that is top ranked in NIST.
- Look for additional features like passive liveness detection.
- Compare performance with slightly skewed, blurry, or dark images.
- Ask the service provider if it is possible to integrate with other third party systems you use, how easy is this, and if any kind of coding is required.
- Compare solutions from several service providers, or just opt for HyperVerge!
Benefits of HyperVerge
HyperVerge face recognition is powered by an AI with industry leading precision and recall. There are several reasons why HyperVerge is leading the pack.
- The uptime is nearly 100%. This means that customer drop offs do not happen due to non availability of the AI. Customers experience seamless onboarding too without any hiccups.
- The network latency is also low. This means that requests are attended to with great speed and the response time is very good.
- Real-time feedback spots errors immediately, giving the option for the user to resubmit details and ensuring an error-free, smoother onboarding.
- Low bandwidth and poor quality of images also do not detrimentally affect the accuracy of the AI as the AI is well trained on a number of samples and can course correct for these problems.
- It supports a number of devices on both iOS and Android, the two most popular mobile platforms in the world.
- It has been trained on people of different ethnicities across regions in several countries spread across the globe. This also makes it racially diverse and less susceptible to any kind of bias.
When you pick face recognition software, do not compromise on the AI accuracy levels or be misinformed as to how well it performs. Without accurate AI/ML, you cannot expect consistent results in face recognition. Talk to us today to take your business to the next level with HyperVerge AI-powered face recognition and OCR.