Few-shot image steganalysis for steganographic algorithm misalignment
In practical steganalysis applications,the steganographic algorithms under test are often un-known,making it challenging to obtain a sufficient number of labeled samples,leading to stegano-graphic algorithm misalignment issues. To enhance the detection performance of steganalysis when the steganographic algorithm is unknown and only a limited number of labeled images are available,a new steganalysis network,BTONet,is proposed. Firstly,an enhanced SRNet incorporating a bottleneck attention mechanism,referred to as BAMSRNet,serves as the feature extraction module of BTONet. BAMSRNet focuses on the texture area in both spatial and channel dimensions,addressing the limita-tions of directly using SRNet in scenarios with a limited number of labeled images to extract discrimi-native features. Secondly,an integration of orthogonal projection loss and cross-entropy loss is pro-posed to bolster the orthogonality between different categories from the perspectives of features and predicted labels,thereby enhancing the performance of the classification module. Finally,BTONet is evaluated against four well-established spatial domain deep steganalysis algorithms in terms of detec-tion accuracy,training time,testing time,and algorithm stability under steganographic algorithm mis-alignment conditions,alongside conducting ablation experiments. Experimental results indicate that,compared to current state-of-the-art deep learning-based steganalysis approaches,BTONet achieves superior detection performance in few-shot learning scenarios,with detection performance improve-ments ranging from 1.02％ to 10.35％. Moreover,BTONet demonstrates exceptional stability by re-ducing the variance in detection accuracy to 1/60 to 1/20 of the compared steganography algorithms.
万方数据
维普
