DASM: Domain-Aware Sharpness Minimization for Multi-Domain Voice Stream Steganalysis

2026-05-19

AI Summary

This work addresses the limited cross-domain generalization of existing speech stream steganalysis methods. To this end, it introduces domain-aware sharpness minimization into this task for the first time. Through Hessian-based analysis, the study reveals that conventional models tend to converge to sharp minima, which harms generalization. The authors propose a novel optimizer that integrates domain-supervised contrastive learning with a sharpness-aware mechanism, dynamically adjusting inter-domain loss weights to guide optimization toward flatter minima. Coupled with an adaptive domain gap modulation strategy, the proposed approach substantially enhances cross-domain generalization performance. Extensive experiments demonstrate that it significantly outperforms current state-of-the-art methods in multi-domain speech steganalysis, exhibiting superior robustness and generalization capability.

Abstract

The growing use of information hiding in network streaming media for covert communication poses a significant security threat, necessitating the development of robust detection technologies. However, existing steganalysis methods for network voice streams mostly rely on data distributions in specific scenarios, making it difficult to adapt to the practical detection needs of non-homologous data distributions. Through Hessian analysis, we find that the loss landscapes of mainstream models are dominated by numerous saddle points and sharp local minima, rendering them highly sensitive to data distribution shifts and fundamentally limiting generalization. Therefore, we propose a new optimizer, Domain-Aware Sharpness Minimization (DASM). The core mechanisms of DASM consist of two aspects: first, it integrates domain-supervised contrastive learning with sharpness-aware optimization, explicitly preserving inter-domain feature separation while seeking flat minima; second, we design an adaptive domain gap modulation strategy that dynamically calibrates the optimization loss weights by sensing the real-time feature separability of different domains. Extensive experimental results demonstrate that our method outperforms the state-of-the-art methods by a large margin and achieves excellent generalization and robustness.

Problem

• steganalysis
• multi-domain
• generalization
• data distribution shift
• voice stream

Innovation

• sharpness minimization
• domain-aware optimization
• contrastive learning
• loss landscape
• steganalysis