People in Photo Albums (PIPA) dataset

People in Photo Albums (PIPA) is a dataset of facial photographs intended to recognise peoples' identities in photo albums in an unconstrained setting.

Created by Facebook and UC Berkeley and published in 2015, the dataset comprises 60,000 facial images of approximately 2,000 people, of which 32,518 photographs were downloaded from Flickr.

Most of the photos are semi-public images of children, family dinners, weddings, and other personal events. 

Risks and harms 🛑

The People in Photo Albums (PIPA) dataset raises privacy concerns by using personal photos without explicit consent, potentially exposing individuals' relationships, activities, and identities to unintended analysis and compromising their privacy in social contexts.

Transparency and accountability 🙈

The People in Photo Albums dataset is seen to suffer from several transparency limitations.

• Lack of consent. The dataset contains 60,000 face images of about 2,000 individuals, with 32,518 photos taken from Flickr without obtaining explicit consent from the individuals depicted.

• Data collection. While some information is provided about using Flickr images, the full extent of the data collection and curation process is not entirely transparent.

• Potential misuse. The dataset has been used for purposes beyond its original intent, including military research and surveillance applications, raising ethical concerns about unexpected uses of personal data.

• Limited demographic information. There is a lack of comprehensive information about the subjects' demographics, making it difficult to assess potential biases or representativeness.

• Restricted access. The original dataset is no longer distributed by Berkeley, though it remains available through other sources, potentially limiting the ability of researchers to independently verify or analyse its contents.

• Usage guidelines. There appears to be a lack of clear guidelines or restrictions on how the dataset can be used, potentially leading to misuse or unethical applications of the biometric data.

• Incomplete metadata. The dataset includes tags that suggest the presence of sensitive information (e.g., DoD, Military), but there's no clear explanation of how this metadata was collected or verified. 

• Unclear image resolution. It is not specified which resolution of Flickr images was used in the dataset, potentially affecting the quality and consistency of the data.

Research, advocacy 🧮

• Harvey, A., LaPlace, J. (2021). Exposing.ai

• Zhao J., Li J., Cheng Y., Zhou L., Sim T.,Yan S., Feng J. (2018). Understanding Humans in Crowded Scenes: Deep Nested Adversarial Learning and A New Benchmark for Multi-Human Parsing