AI Powered Ads: Foolproof and Fair?

Description

We are currently living in a time where AI technology is becoming more and more integrated with our everyday lives. This is the case of Westfield's use of Quividi SmartScreen Technology, which uses facial analysis to change advertising messages based on shopper's gender and age. Less intrusive applications like these establish social license for similar technology to be used in more serious situations like health and security. Thus, it is vital we be robust and thorough in our evaluation of these facial analysis systems. Our project investigated two questions: are these models fair in the New Zealand context? Are they foolproof? The model evaluated is Deepface, a 'state-of-the-art' model which boasts a high accuracy, which is capable of classifying a person's gender, ethnicity, age, and emotion. For the 'fair' question, we created our own New Zealand specific database and tested Deepface using a variety of statistical methods. We found that Maori are treated as White 55% of the time, and Pasifika as equally split between Asian, Latino, and Black. We also found evidence the model is biased towards classifying male, white, young adult (20-29) and happy faces. For the 'foolproof' question, we successfully trained adversarial attacks in the form of colour perturbartion within a constrained accessory area. We made attacks for dodging a person's true gender, ethnicity, and emotion, or impersonating a specific gender, ethnicity, or emotion. We tested these attacks digitally and found convincing success (>90%), including for Maori and Pasifika. Then, we printed these accessories and did rounds of physical testing, iteratively improving on physical attack performance to develop a set of accessories that do reasonably well in real life.