Exploring the Effect of Visual-Based Subliminal Persuasion in Public Speeches Using Explainable AI Techniques

Klaus Weber, Lukas Tinnes, Tobias Huber, Elisabeth André

Abstract: When it comes to persuading other people, non-verbal cues play an important role in order to be successful. Mostly, people use these non-verbal cues subconsciously and, from the perspective of the persuadee, are not aware of the subliminal impact of them. To raise awareness of subliminal persuasion, we analyzed videos of different political public speeches. We used the labels of three annotators to train three subjective neural networks capable of predicting their degree of perceived persuasiveness based on the images as input only. We then created visualizations of the predictions for each network/annotator to draw conclusions about what the annotators have most likely focused on. For that, we employed layer-wise relevance propagation (LRP) that highlights the most relevant image sections for each prediction. Our results show that techniques like LRP can help uncover existing subliminal bias.