Smartphone Camera Identification from Low-Mid Frequency DCT Coefficients of Dark Images

Area of Science:

• Digital Forensics
• Image Processing
• Machine Learning

Background:

• Camera sensor identification is crucial for digital forensics and authentication.
• Previous methods often focus on different image properties or channels.
• Dark Signal Nonuniformity (DSNU) presents a viable characteristic for sensor fingerprinting.

Purpose of the Study:

• To develop and evaluate a methodology for smartphone camera sensor identification.
• To leverage Dark Signal Nonuniformity (DSNU) properties for unique sensor fingerprinting.
• To compare the effectiveness of different machine learning models and image channels for classification.

Main Methods:

• Acquisition of dark images by covering the smartphone camera lens.
• Extraction of low and mid-frequency AC coefficients from the Discrete Cosine Transform (DCT) of dark images.
• Classification of extracted features using machine learning algorithms, including K-Nearest Neighbor (KNN) and a wide neural network.

Main Results:

• The wide neural network achieved superior classification performance compared to traditional algorithms like KNN and a more complex network.
• Analysis revealed that the blue image channel offered the best discriminatory power for sensor identification.
• This finding contrasts with prior research that suggested the green channel is optimal.

Conclusions:

• Dark Signal Nonuniformity (DSNU) is an effective characteristic for smartphone camera sensor identification.
• A simple, wide neural network model demonstrates high efficacy in classifying DSNU patterns.
• The blue channel's superiority in this context warrants further investigation for sensor forensics applications.