CIOs Start To Deal With The Limitations Of Image-Recognition Filters

It turns out that there are ways to get around image-recognition filters
It turns out that there are ways to get around image-recognition filters
Image Credit: jane cornwell

As your company’s CIO it is your responsibility to understand the importance of information technology and to use it to secure the company’s assets. What this means on a daily basis is that you are probably installing firewalls and creating white lists for who can access what applications and servers. However, as life become more and more complicated, you may also be dealing with the arrival of image-recognition filters for everything from screening online content and people arriving at your company’s buildings. However, it turns out that these new AI driven systems can be fooled…

How To Fool An Image-Recognition Filter

So before we dive too deeply into this topic, let’s first deal with the question of why someone would want to fool an image-recognition filter? It turns out that there are lot of people out there who would like to find a way around any image-recognition system that you put into place. The most obvious of these people are the ones who want to sneak into your buildings or into your secured server rooms. Less obvious, but still of concern to you are the people who want to add things to your websites that you don’t want them to put there. The most obvious of these are images having to do with guns, violence, drugs, or sex.

The ability to fool an image-recognition system poses a big risk to any firm that is making use of such systems. Image-recognition systems are being used by the big social media firms to screen content. They are also being used by security systems, self-driving cars, and numerous other systems. The ability to fool systems like this points out how hard it is to prevent image-recognition systems from being either fooled or gamed.

There are different names for what attackers are doing in order to try to get past image-recognition systems. When a type of digital camouflage is placed over a picture in order to get it past a filter, this is called an “adversarial attack”. The goal is to leave a picture looking normal to a human eye, but get it past image-recognition software. Facebook has been struggling with having its users upload content that they have banned. Recently a video of a tragedy kept getting uploaded and because it had be altered, it was able to get past Facebooks filters. What this meant is that the digital fingerprint that Facebook had assigned to the original video could no longer be relied on to detect copies.

Next Steps In Creating Better Image-Recognition Filters

Over at Facebook, they use artificial intelligence (AI) software to scan content that users have uploaded in order to detect hate speech, terrorist propaganda, and spoofed accounts. However, because they realize that the software is not perfect, they have also hired 30,000 human content moderators to scan and remove inappropriate content. The Facebook software has the ability to detect faces, objects, and types of behavior.

Google is dealing with the same image-recognition issues that Facebook is dealing with. Google’s goal is to keep harmful content off of its YouTube property. Google also relies on AI software to detect inappropriate content. However, they realize that software is not able to do a complete job and so they have hired 10,000 people to act as content police. Google’s goal is to improve their AI software and decrease the number of humans that they have to employ to perform this task. The bad guys have started to take banned videos and create new versions of them that are slightly fuzzy in order to get them to slide past image-recognition systems.

The people who are responsible for creating the software that performs the functions of image-recognition filters realize that they still have a way to go. Most of these firms are open to having people provide them with examples where their software failed to detect banned content. This helps them to conceive defense against unintended uses of the AI models that are part of their products. One of the most troubling spoofs of an image-recognition system happens when individuals are able to hold up pictures in front of their face and this causes them to not be detected by the software system.

What All Of This Means For You

The person who has the CIO job has a lot of different responsibilities. Among these responsibilities is the obligation to secure the company’s buildings and to prevent harmful content from being loaded onto company websites and servers. This is a complex task that has to be performed all the time. The person in the CIO position really only has one choice, they have to implement an image-recognition filter system. However, CIOs need to understand that these systems come with their own set of limitations.

People might be interested in trying to fool an image-recognition system if they wanted to gain access to some location where they were not permitted or if they wanted to upload content such as images or video that the company banned. Fooling an image-recognition system is a big deal because they are being used in so many different ways. Facebook is one company that has to deal with trying to use image-recognition systems to prevent harmful content from being uploaded. Facebook has hired 30,000 workers to help its software detect inappropriate content. Google has hired 10,000 workers to do the same thing. The people who make the image-recognition systems understand that their systems are not perfect and so they are constantly trying to improve them.

Image-recognition systems are becoming an important part of how companies are run. As the CIO is it your responsibility to make sure that your company has one of these systems implemented and that it operates correctly. You need to understand that today’s image-recognition systems have limitations and so they are not perfect. You’ll need to have a backup plan until the systems improve in their ability to catch all of the inappropriate content. The systems will keep getting better, we just have to help them along until they can do the job all by themselves.

– Dr. Jim Anderson Blue Elephant Consulting –
Your Source For Real World IT Department Leadership Skills™

Question For You: What do you think is the best way to train an image-recognition system so that it can detect inappropriate content?

Click here to get automatic updates when The Accidental Successful CIO Blog is updated.
P.S.: Free subscriptions to The Accidental Successful CIO Newsletter are now available. Learn what you need to know to do the job. Subscribe now: Click Here!

What We’ll Be Talking About Next Time

You would think that being a CIO at a rental car company would be a pretty straight forward job. I mean, if you understand the importance of information technology and can keep track of the employees and the cars then you are pretty much on top of your job. However, the way that we use cars is getting ready to undergo some significant changes with the arrival of Uber and ride sharing services. How is the CIO of Enterprise Holdings going to help the company to evolve to meet a changing world?