Covert photo classification by fusing image features and visual attributes.

Area of Science:

• Computer Science
• Information Security
• Digital Forensics

Background:

• Covert photos pose significant privacy risks due to their concealed acquisition.
• Automatic identification is crucial for privacy protection but challenging due to semantic similarity and data scarcity.
• Existing methods lack effectiveness in distinguishing covert imagery.

Purpose of the Study:

• To address the challenge of classifying covert photos.
• To develop an effective automatic system for identifying secretly captured images.
• To enhance digital privacy through automated detection of invasive photography.

Main Methods:

• Collected and verified a large dataset of 2500 covert photos.
• Conducted a user study to establish a human performance baseline for covert photo identification.
• Developed a novel covert photo classification algorithm using multiple kernel learning to fuse diverse image features and visual attributes.

Main Results:

• Achieved an average classification rate (1-EER) of 0.8940 on the collected dataset.
• The proposed algorithm significantly outperformed other state-of-the-art image classifiers.
• The developed method also surpassed human performance in distinguishing covert photos.

Conclusions:

• The novel multiple kernel learning approach effectively classifies covert photos.
• The developed dataset and user study provide valuable resources for future research in this area.
• This work represents a significant advancement in automated detection of privacy-invasive imagery.