Year: 2025
Type: conference-paper
Abstract: Post-hoc calibration is an important methodology for im-proving the reliability of confidence estimates in deep neu-ral networks (DNNs). While modern DNNs achieve state-of-the-art performance across various domains, the output they provide often fails to align with the true likelihood of their predictions, a phenomenon known as miscalibration. This misalignment poses challenges for tasks like uncertainty quantification, score level fusion, and threshold se-lection in the case of binary detection systems. Furthermore post-hoc calibration methods, such as temperature scaling, Platt scaling, and isotonic regression [17] provide practical solutions to alleviate miscalibration, without retraining. In this work, we propose the use of post-hoc calibration for multimedia forensics applications, by focusing on the detection of synthetically AI generated images. In particular, we show how the application of well known post-hoc calibration methodologies can help to improve the interpretability of AI generated images in terms of likelihood ratios, and can also help to adjust the detection threshold in the presence of different AI generators, considered in the training and calibration sets.
Institution: University of Siena
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4409917780