Quantitative Metrics for Evaluating Explanations of Video DeepFake Detectors

TL;DR

This paper introduces quantitative metrics to evaluate the quality and informativeness of explanations for video DeepFake detectors, addressing a gap in explainability assessment for AI models in content moderation.

Contribution

It proposes a simple set of metrics for human-centric evaluation of explanations, enabling better comparison and understanding of explanation methods for DeepFake detectors.

Findings

Metrics effectively differentiate explanation quality

Improved explanation methods enhance interpretability

Evaluation impacts model deployment strategies

Abstract

The proliferation of DeepFake technology is a rising challenge in today's society, owing to more powerful and accessible generation methods. To counter this, the research community has developed detectors of ever-increasing accuracy. However, the ability to explain the decisions of such models to users is lacking behind and is considered an accessory in large-scale benchmarks, despite being a crucial requirement for the correct deployment of automated tools for content moderation. We attribute the issue to the reliance on qualitative comparisons and the lack of established metrics. We describe a simple set of metrics to evaluate the visual quality and informativeness of explanations of video DeepFake classifiers from a human-centric perspective. With these metrics, we compare common approaches to improve explanation quality and discuss their effect on both classification and explanation performance on the recent DFDC and DFD datasets.