More than Meets the Eye: Understanding the Effect of Individual Objects on Perceived Visual Privacy

TL;DR

The paper addresses the problem that privacy research often treats an image as a whole, ignoring how individual objects contribute to perceived privacy. It adopts a mixed-methods approach with synthetic stimuli and 92 participants to study foreground-background object interactions across two contexts, revealing how background PSOs, environment, and co-presence modulate privacy judgments. Key contributions include empirical evidence of foreground-background interactions, dominance and co-presence effects, and qualitative insights into users' mental models and risk inferences. The work advances context-aware privacy design, offering guidelines for automated highlighting and selective obfuscation to balance privacy with image utility in social platforms and AI applications.

Abstract

User-generated content, such as photos, comprises the majority of online media content and drives engagement due to the human ability to process visual information quickly. Consequently, many online platforms are designed for sharing visual content, with billions of photos posted daily. However, photos often reveal more than they intended through visible and contextual cues, leading to privacy risks. Previous studies typically treat privacy as a property of the entire image, overlooking individual objects that may carry varying privacy risks and influence how users perceive it. We address this gap with a mixed-methods study (n = 92) to understand how users evaluate the privacy of images containing multiple sensitive objects. Our results reveal mental models and nuanced patterns that uncover how granular details, such as photo-capturing context and copresence of other objects, affect privacy perceptions. These novel insights could enable personalized, context-aware privacy protection designs on social media and future technologies.

Figures (8)

• Figure 1: An example demonstrating the amount of information revealed (intentionally or inadvertently) through a single visual content. The image is sampled from the https://cocodataset.org/#explore and the text descriptions are a combined interpretation by a human observer (one of the authors of this work) and commercial GenAI methods (link to actual image is http://farm3.staticflickr.com/2596/3855197434_72ccaa2ed3_z.jpg).
• Figure 2: Illustrative examples of image generation in café (top) and office (bottom) environments. Each sequence progresses from a face only foreground PSO to combinations with additional background PSOs.
• Figure 3: Average comfortability level ratings for foreground PSOs and the effect of a single background PSO.
• Figure 4: Average comfortability level ratings for each foreground PSO, showing the baseline, the effect of the two background PSO most often rated as having the lowest effect when alone (see Table \ref{['tab:least_private_counts']}), and the effect of co-presence with both background PSOs present in the image (both).
• Figure 5: Visual accompanying Q12. Scale: 1 = Not comfortable at all, 2 = Slightly uncomfortable, 3 = Neutral, 4 = Slightly comfortable, 5 = Very comfortable