Steganographer detection via a similarity accumulation graph convolutional network.

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Steganographer detection is crucial for identifying covert communication in social networks.
• Existing methods often overlook the relationships between image features and user representations.
• Current approaches struggle to differentiate subtle nuances between guilty and innocent users.

Purpose of the Study:

• To formulate steganographer detection as a multiple-instance learning problem.
• To propose a novel similarity accumulation graph convolutional network (SAGCN) for enhanced user representation.
• To improve the accuracy and reliability of detecting users engaged in covert image sharing.

Main Methods:

• Formulated steganographer detection as a multiple-instance learning (MIL) problem, with users as bags and images as instances.
• Developed a SAGCN to represent users as weighted graphs based on image feature similarity.
• Implemented a novel graph reconstruction and pooling strategy to address oversmoothing and enhance node distinction.

Main Results:

• The proposed SAGCN framework demonstrated superior effectiveness and reliability compared to state-of-the-art and graph-based models.
• The method showed strong performance across diverse image domains and large-scale social media scenarios.
• The network's generalizability to other MIL problems was also indicated.

Conclusions:

• The SAGCN offers a robust approach to steganographer detection by leveraging inter-instance relationships.
• This method effectively captures user-level patterns for identifying covert communication.
• The proposed framework advances the field of steganographer detection and has broader applicability in MIL tasks.