Lightweight image steganalysis with block-wise pruning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep learning models for image steganalysis achieve high detection performance but are computationally expensive.
• Existing models are often too heavy and costly for practical deployment in real-world applications.
• There is a need for lightweight and efficient steganalysis models.

Purpose of the Study:

• To develop an effective model design strategy for lightweight image steganalysis.
• To reduce the computational cost and size of deep steganalysis models.
• To maintain high detection performance with a simplified model architecture.

Main Methods:

• Proposing a block removal strategy for deep classification networks in steganalysis.
• Gradually removing convolutional neural network blocks from deeper layers.
• Evaluating model performance on BOSSBase and ALASKA#2 datasets.

Main Results:

• The proposed block removal strategy significantly reduces model size and FLOPs (Floating Point Operations Per Second).
• EfficientNet-B0 variants achieved 9.58% smaller size and 2.16% fewer FLOPs compared to the baseline.
• Detection accuracy remained high, on par with the baseline at 90.73% and 82.40% on respective datasets.
• Analyses indicated that early layers of the network are crucial for effective steganalysis.

Conclusions:

• A simple block removal strategy can create lightweight yet effective image steganalysis models.
• The developed models are suitable for practical deployment due to reduced computational requirements.
• Early network layers contain essential features for robust image steganalysis.