Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification

TL;DR

This work addresses source camera identification (SCI), a problem crucial for verifying image provenance amid metadata removal and AI-generated manipulation. It systematically compares three SCI approaches—JPEG artifact analysis, PRNU-based sensor fingerprinting, and CNN-based end-to-end classification—on a four-device smartphone dataset. The key finding is that traditional, interpretable features, particularly JPEG-AC statistics with a linear SVM, outperform a CNN baseline, achieving $Acc_ ext{JPEG}=0.90$, $Acc_ ext{PRNU}=0.71$, and $Acc_ ext{CNN}=0.29$. The results motivate hybrid approaches and larger datasets to enhance robustness for real-world multimedia security and forensics applications.

Abstract

One of the most important tasks in computer vision is identifying the device using which the image was taken, useful for facilitating further comprehensive analysis of the image. This paper presents comparative analysis of three techniques used in source camera identification (SCI): Photo Response Non-Uniformity (PRNU), JPEG compression artifact analysis, and convolutional neural networks (CNNs). It evaluates each method in terms of device classification accuracy. Furthermore, the research discusses the possible scientific development needed for the implementation of the methods in real-life scenarios.

Figures (4)

• Figure 1: Overall accuracy comparison of the JPEG-based, PRNU-based, and CNN-based classification methods.
• Figure 2: Confusion matrix for the JPEG artifact-based classifier.
• Figure 3: Confusion matrix for the PRNU-based classifier.
• Figure 4: Confusion matrix for the CNN classifier.