PEAK is a privacy assistant that explains privacy labels by generating human-understandable explanations. PEAK is composed of the following five stages:
- [Annotated Data] The starting point for PEAK is a set of labeled public and private images. The data could come from a user's own history of personal online images (i.e. both shared and not shared), or from a publicly available dataset of labeled images. You can also download images here in their original "jpg" format.
- [Generate Tags] The second stage involves assigning tags to each image. The generated tags are in plain language such as "tree" or "baby", and they can be provided by users themselves or automatically generated by a tool such as Clarifai.
- [Extract Topics] The third stage is using the set of labeled and tagged images to perform topic modelling, which is a technique used to extract latent topics from textual information (i.e. the tags).
- [Classify Images] The fourth stage is to train a tree-based ML algorithm for the binary image classification. Additionally, the contributions of each topic to the privacy decision (i.e. positive or negative effect) are computed.
- [Generate Explanations] The final stage is identifying explanation categories for images, which essentially are image profiles with certain characteristics in terms of topic contributions. For instance, there may be a single dominant topic that pushes the prediction overwhelmingly in one direction. Based on the topics' contributions to privacy prediction, images are assigned to one of four explanation categories we identified: Dominant, Opposing, Collaborative, and Weak. Finally, for each explanation category, there is a distinct textual and visual explanation pattern based on the image's topics and the relationships between topics.
You can read&cite our papers: Can We Explain Privacy? and PEAK.
If you have any questions or comments, please feel free to leave them as a new issue or email.