Deep Learning Image Age Approximation - What is More Relevant: Image Content or Age Information?

The field of temporal image forensics is the science of approximating the age of a digital image relative to images from the same device. For this purpose, classical methods exist that exploit handcrafted features based on hidden age traces (i.e., in-field sensor defects). In contrast to these classical methods, a Convolutional Neural Network (CNN) learns the features used independently. This has the benefit that other (unknown) age traces can be exploited. However, this also carries the risk of learning non-age-related features to predict the age class. In this work, we analyze the features learned by a standard CNN trained in the context of image age approximation on regular scene images. To analyze whether the model learned to exploit hidden age traces or just other general, non-age-related intra-class properties (e.g., common scene properties or lighting conditions), we applied methods from the field of Explainable Artificial Intelligence (XAI). This analysis is performed with 14 models trained with images from 14 different devices from two datasets.