Attack-Aware Deepfake Detection with Robustness and Calibration

This paper addresses the critical problem of deepfake detection, focusing on robustness against counter-forensic manipulations. It proposes a novel architecture combining red-team training and randomized test-time defense, aiming for well-calibrated probabilities and transparent evidence. The approach is particularly relevant given the evolving sophistication of deepfake generation and the need for reliable detection in real-world scenarios. The focus on practical deployment conditions, including low-light and heavily compressed surveillance data, is a significant strength.