Source-linked AI summary
Yedrouj-Net: An efficient CNN for spatial steganalysis
Mehdi Yedroudj, Frederic Comby, Marc Chaumont
TL;DR
Spatial steganalysis needed stronger evidence that CNNs could surpass classical rich-model ensembles without relying on increasingly elaborate designs. The paper proposes Yedroudj-Net, combining established CNN components and augmented training data, and reports a 16% error-probability decrease versus RM+EC on WOW at 0.2 bpp.
Problem
Deep-learning steganalysis had produced results close to classical approaches, but had not delivered the larger improvements observed in other image-processing tasks.
Method
The paper designs Yedroudj-Net by combining a fixed high-pass filterbank, Truncation activations, five convolutional blocks with Batch Normalization and Scale, and a fully connected classifier, using augmented training data.
Results
16% error probability decrease was obtained for WOW at 0.2 bpp versus RM+EC, after adding BOWS2 and virtually augmenting the learning database.
Takeaways & Limitations
Yedroudj-Net achieved the best reported results in the augmented comparison, while larger databases further improved CNN steganalysis performance.
Takeaways & Limitations
Ye-Net remains sensitive to learning-rate initialization, whereas Batch Normalization makes Yedroudj-Net and Xu-Net less sensitive to this setting.
Abstract
from arXiv · showhide
For about 10 years, detecting the presence of a secret message hidden in an image was performed with an Ensemble Classifier trained with Rich features. In recent years, studies such as Xu et al. have indicated that well-designed convolutional Neural Networks (CNN) can achieve comparable performance to the two-step machine learning approaches. In this paper, we propose a CNN that outperforms the state-ofthe-art in terms of error probability. The proposition is in the continuity of what has been recently proposed and it is a clever fusion of important bricks used in various papers. Among the essential parts of the CNN, one can cite the use of a pre-processing filterbank and a Truncation activation function, five convolutional layers with a Batch Normalization associated with a Scale Layer, as well as the use of a sufficiently sized fully connected section. An augmented database has also been used to improve the training of the CNN. Our CNN was experimentally evaluated against S-UNIWARD and WOW embedding algorithms and its performances were compared with those of three other methods: an Ensemble Classifier plus a Rich Model, and two other CNN steganalyzers.
1. INTRODUCTION
Deep-learning steganalysis progressed from auto-encoders and CNNs toward architectures approaching classical methods, but improvements over rich-model ensembles remained limited. Yedroudj-Net addresses this gap with a single CNN designed to improve spatial-steganalysis performance without ensembles or task-specific design.
- Prior work: Early deep-learning steganalysis used auto-encoders, followed by CNNs and CNN ensembles that approached state-of-the-art performance.Xu-Net was used as the base learner for one CNN ensemble.
- Motivation: Deep-learning methods had not achieved the roughly 10% improvement seen in other image-processing tasks over classical Ensemble Classifier and Rich Model approaches.The comparison concerns classical approaches using an Ensemble Classifier with Rich Models or selection-channel awareness.
- Motivation: Recent CNN improvements commonly relied on ensembles, Rich-Model-inspired topology changes, or ResNet, often requiring substantial design or experimental effort for small gains.These trends were identified for steganalysis CNNs in 2017.
- Contribution: Yedroudj-Net is a single CNN for spatial steganalysis designed to outperform the state of the art without image-domain-specific design, CNN ensembling, transfer learning, or virtual database augmentation.The paper compares it with Xu-Net, Ye-Net, and an Ensemble Classifier using Spatial-Rich-Models.
- Contribution: The proposed network is reported to be insensitive to hyperparameter initialization and to converge easily.The paper states that this behavior is discussed later in Section 3.
2. YEDROUDJ-NET
Yedroudj-Net combines fixed residual preprocessing, five convolutional blocks, and a three-layer classification module to distinguish cover from stego images. Its design uses specialized activations, normalization, pooling, and fully connected layers to represent and classify weak steganographic signals.
- Architecture: The CNN comprises a preprocessing block, five convolutional blocks, and three fully connected layers followed by a softmax over cover and stego labels.The classification module uses 256, 1024, and 2 neurons in its successive fully connected layers.
- Preprocessing: A predefined bank of 30 SRM high-pass filters extracts noise residuals before feature learning, with fixed 5×5 kernels integrated into the CNN.The preprocessing weights are not optimized during training.
- Convolutional module: The convolutional blocks use 5×5 kernels in Blocks 1–2 and 3×3 kernels in Blocks 3–5, without convolutional biases because bias terms move to the Scale Layer.This kernel arrangement follows the stated design for the five-block convolutional module.
- Activations: Block 1 applies absolute value activation, Blocks 1–2 use truncation, and Blocks 3–5 use ReLU to process residual features.Absolute value enforces sign symmetry, truncation limits large values, and ReLU is selected for performance and fast gradient computation.
- Normalization: Batch normalization with a Scale Layer normalizes and then adjusts feature distributions, while supporting larger learning rates and reducing sensitivity to initialization.The cited design contrasts its BN-plus-Scale arrangement with Xu-Net.
- Pooling: Average pooling in Blocks 2–5 downsamples feature maps, while global average pooling in the final block removes location information about embedded pixels.The first block has no pooling to avoid information loss.
3. EXPERIMENTS
The experiments compare Yedroudj-Net with CNN and handcrafted baselines under controlled spatial-steganalysis settings, then evaluate database augmentation. Yedroudj-Net achieves the strongest reported error probabilities, while larger training bases further improve CNN performance.
- CNN comparisons: Yedroudj-Net combines five convolutional layers, Batch Normalization, Absolute Value activation, a 30-filter preprocessing bank, Truncation in Blocks 1 and 2, and three fully connected layers.Compared with Xu-Net and Ye-Net, the architecture differs in preprocessing, activation choices, normalization, convolutional depth, and fully connected capacity.
- Experimental setup: The evaluation compares Yedroudj-Net, Xu-Net, Ye-Net, and SRM+EC on WOW and S-UNIWARD at 0.2 and 0.4 bpp.All methods use the same subsampled BOSSBase images; CNN experiments use 256×256 images and separate training, validation, and untouched testing sets.
- General performance comparisons: For WOW, Yedroudj-Net lowers error probability versus SRM+EC by 8% at 0.2 bpp and 11% at 0.4 bpp.For S-UNIWARD, it matches SRM+EC at 0.2 bpp and is 2% lower at 0.4 bpp.
- General performance comparisons: Yedroudj-Net is 2% to 6% better than Xu-Net and 3% to 9% better than Ye-Net across the two embedding algorithms and payloads.The other CNNs are not always superior to SRM+EC without an ensemble or increased learning database.
- Base augmentation: Increasing the training base improves all three CNNs on WOW at 0.2 bpp, with Yedroudj-Net reducing detection error probability by 4% using BOSS+BOWS2.With virtual augmentation, Yedroudj-Net reaches an error probability of 20.8%, a 16% improvement over RM+EC at 36.5% trained on BOSSBase.
- Base augmentation: Even 112 000 virtually augmented cover/stego pairs are insufficient for CNNs with 5–7 blocks, although larger databases improve performance as convergence time increases.Yedroudj-Net has approximately 500 thousand parameters, compared with approximately 50 thousand for Xu-Net.
4. CONCLUSION
Yedroudj-Net combines several design choices into a CNN for spatial steganalysis that beats state-of-the-art approaches, with further gains from enlarging and augmenting the learning database.
- Yedroudj-Net is a simple CNN for spatial steganalysis that beats state-of-the-art approaches in a clairvoyant scenario without selection-channel knowledge.
- Its performance improvement combines a preprocessing filter bank, Truncation activation, and Batch Normalization associated with a Scale Layer.
- 16% error-probability decrease was obtained on WOW at 0.2 bpp compared to RM+EC after adding BOWS2 and virtually augmenting the learning database.