Is it the model or the metric -- On robustness measures of deeplearning models

View on arXiv ← Back to list

Authors: Zhijin Lyu, Yutong Jin, Sneha Das

Published: 2024-12-13 02:26:58+00:00

AI Summary

This paper investigates the robustness of deepfake detection models by introducing a new metric, Robust Ratio (RR), which complements the existing Robust Accuracy (RA). The authors demonstrate that while models may exhibit similar RA, their RR values vary significantly under different perturbation levels, revealing disparities in their robustness.

Abstract

Determining the robustness of deep learning models is an established and ongoing challenge within automated decision-making systems. With the advent and success of techniques that enable advanced deep learning (DL), these models are being used in widespread applications, including high-stake ones like healthcare, education, border-control. Therefore, it is critical to understand the limitations of these models and predict their regions of failures, in order to create the necessary guardrails for their successful and safe deployment. In this work, we revisit robustness, specifically investigating the sufficiency of robust accuracy (RA), within the context of deepfake detection. We present robust ratio (RR) as a complementary metric, that can quantify the changes to the normalized or probability outcomes under input perturbation. We present a comparison of RA and RR and demonstrate that despite similar RA between models, the models show varying RR under different tolerance (perturbation) levels.

Key findings

While Robust Accuracy (RA) showed similar performance across models, Robust Ratio (RR) revealed significant variations in model robustness under different perturbation levels. RR exhibited a more linear relationship with tolerance for video datasets compared to image datasets, possibly due to the temporal nature of video data and the way adversarial attacks are applied.

Approach

The paper proposes a new robustness metric, Robust Ratio (RR), which quantifies changes in model output probabilities under input perturbations. It compares RR with the existing Robust Accuracy (RA) metric to assess the robustness of deepfake detection models against various adversarial attacks.

Datasets

UADFV, FF++, Celeb-DF-v2

Model(s)

Meso4, Meso4Inception, ResNet34

Author countries

Denmark, UK

← Previous