Thesis (Ph.D.) - Indiana University, School of Informatics, Computing, and Engineering, 2020Billions of photos are being shared on social media platforms every day. A large portion of these photos are taken in public places, and may contain people who were inadvertently captured (i.e., bystanders) and are not important for the subject matter of the photos. When these photos are shared online, they reveal the bystanders' identity, location, and other privacy-sensitive information to a potentially unbounded number of internet users. Social media users not only share photos they own but also re-share photos from their peers and those they find on the internet; for example, the sharing of image macros or memes on social media has risen in popularity. Internet users create memes using photos of other people who are often unknown to them. Such photos usually portray people in embarrassing situations, which are highlighted and ampli fied with additional text or captions. These photos can go `viral' and cause severe personal, social, and professional consequences to the photo subjects. While the research community has made signi cant efforts to reduce photosharers' privacy risks on social media, protecting the privacy of people who do not actively take part in photo-taking or sharing activities, e.g., bystanders and meme subjects, has not received adequate attention. This dissertation proposes machine learning and computer vision-based techniques to reduce bystanders' privacy risks. More specfi cally, we offer an automated and scalable system to detect bystanders in images so that their privacy can be protected by, e.g., removing or obfuscating them using image transforms. In an online study, we evaluated the effectiveness and usability of commonly used image transforms. We constructed and empirically validated models of interactions among image filters and utility variables. Based on these models, we proposed a principled approach to design novel obfuscations to balance the privacy-utility trade-o s. To protect the privacy of meme subjects, we explored the potential of behavioral interventions to discourage meme sharing. Through controlled experiments, we identfi ed demographic factors and personality traits that affect behaviors regarding photo sharing that may threaten other people's privacy. We also discovered links between people's personality traits and their reactions to privacy nudges that were designed to discourage them from sharing memes. These results can be used to develop direct and personalized interventions to stimulate privacy-respecting and prosocial behaviors among social media users.