An image steganalysis method and system based on heterogeneous inverse regularization and specific tensor reconstruction
By employing heterogeneous inverse regularization and specific tensor reconstruction, the problems of feature purity and generalization ability of deep steganalysis models in extremely low signal-to-noise ratio environments are solved, achieving more accurate steganalysis detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-17
AI Technical Summary
Existing deep steganalysis models cannot effectively repair phase breaks in steganalytic residuals, gradient flooding and numerical underflow collapse caused by cross-entropy loss, and spurious associations solidified by traditional knowledge distillation in environments with extremely low signal-to-noise ratios, resulting in insufficient generalization ability.
We employ heterogeneous inverse regularization and specific tensor reconstruction, extract noise residuals through a multi-scale cascaded filter bank, construct a teacher network for feature purification and steganalysis capture, and combine specific tensor cross-reconstruction and heterogeneous inverse gradient truncation fine-tuning to achieve feature purification and adaptive optimization of difficult examples.
It significantly improves feature purity under extremely low signal-to-noise ratios, avoids gradient underflow collapse, and enhances the model's cross-source generalization robustness and detection accuracy.
Smart Images
Figure CN122415489A_ABST