Exploring Biases in Vision Privacy Protection Techniques:

the abstract

In the Spring of 2019, I joined the OU Data Analytics Lab as an undergraduate researcher. Right away, I began working with Dr. Christan Grant and doctoral Jasmine DeHart on a project exploring topics dealing with biases and vision privacy protection techniques.

In recent years machine learning has become a common part of technology and society. Images and can be scanned and classified as a result of object detection tools. Although it has become a large part of the technology we use today, machine learning models are not easy to tailor for a variety of demographics. As a result, issues with bias, fairness, and accountability arise. The primary goal of this study was to understand how biases in computer vision techniques can affect the potential of social media-based privacy leaks.

Throughout the semester, I worked closely with Jasmine to create a computer vision system to identify content that had the potential to leak sensitive information.

what I learned

Prior to doing this study, I had no prior experience with machine learning or image object detection processes. After getting comfortable with python and various libraries, I began working with a popular object detection system called YOLO (You Only Look Once). After developing a list of potential privacy leak categories(Ex: Keys, credit cards, babies, etc.), we began running the YOLO script on our image data sets. We noticed right away, that when detecting images of children, babies of color went undetected frequently. To ensure that the accuracy rate was improved, the visual content collected for the category, ''babies,'' contained a variety of infants ranging in skin tone. Taking all of these factors into account, our original model was revamped to improve the precision and recall of the object detection model. This experience definitely opened my eyes to the importance of representation, even in data sets.


In studies related to image object detection and privacy, the importance of ensuring all populations of every respective group are measured fairly and accurately is imperative to reduce bias in machine learning and computer vision models. For example, on social media networks, computer vision is used to identify individuals who maybe be tagged in other photos.

Working with Jasmine DeHart and Dr. Grant was such a cool experience. And even cooler: our research poster was accepted into the 2019 IEEE Symposium on Security and Privacy.