A Visually Interpretable Forensic Deepfake Detection Tool Using Anchors

Abstract

“Deepfakes” have seen a dramatic rise in recent times and are becoming quite realistic and indistinguishable with the advancement of deepfake generation techniques. Promising strides have been made in the deepfake detection area even though it is a relatively new research domain. Majority of current deepfake detection solutions only classify a video as a deepfake without providing any explanations behind the prediction. However, these works fail in situations where transparency behind a tool’s decision is crucial, especially in a court of law, where digital forensic investigators maybe called to testify if a video is a deepfake with evidence; or where justifications behind tool decisions plays a key role in the jury’s verdict. Explainable AI (XAI) has the power to make deepfake detection more meaningful, as it can effectively help explain why the detection tool classified the video as a deepfake by highlighting forged super-pixels of the video frames. This paper proposes the use of “Anchors” XAI method, a model-agnostic high precision explainer to build the prediction explainer model, that can visually explain the predictions of a deepfake detector model built on top of the EfficientNet architecture. Evaluation results show that Anchors fair better than LIME in terms of producing visually explainable and easily interpretable explanations and produces an anchor affinity score of 70.23%. The deepfake detector model yields an accuracy of 91.92%.