The Impact of Contextual Information in Social Media Posts on Human Deepfake Detection Accuracy
1.09 MB
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Lataukset16
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
Deepfakes are form of synthetic media. They are generated using deep learning algorithms to create realistic but misleading representations of reality. Currently, the deepfake technology is advancing fast and humans are having difficulties separating deepfakes from authentic media, making deepfakes a serious threat. Based on existing research, human deepfake detection accuracy is approximately 60.6%. In addition to low human deepfake detection accuracy, technical deepfake detection solutions have challenges in real-world implementations. Deepfakes are commonly shared and encountered on social media platforms, making these platforms the primary environment for deepfakes. This means that there is a need for human deepfake detection improvement strategies especially on social media as current technical deepfake detection systems are not yet sufficient on these platforms. Existing deepfake detection research is mainly focusing on technical solutions or is not sufficiently considering real-world environment of deepfakes, making research on human deepfake detection in context of social media important.
This thesis examines the impact of contextual information in social media posts on human deepfake detection accuracy and proposes and tests a new improvement strategy called Intent Labeling. It first presents systematic literature review of 88 papers. A survey-based experimental study is then applied to collect quantitative and qualitative data from 73 cybersecurity students. The experimental study has three conditions testing impact of contextual information and Intent Labeling on deepfake detection accuracy. Within-subjects design is applied. In addition to detection accuracy, confidence and maliciousness ratings, contextual cues used in the detection task and opinions on Intent Labeling are collected and analyzed.
Results show that context has impact on detection accuracy. Quantitative data shows that giving suspicious context for image has an effect: detection accuracy increases when deepfake is given with suspicious context and decreases when authentic image is given with suspicious context. Students were more confident when context was given, even though higher confidence was not correlating with higher accuracy. Qualitative data provides insights into what cues students used in the detection task, including visual details, quality of the post, consistency, external knowledge, source, intent and interactions.