With the continuous advancement of steganographic techniques, the task of image steganalysis has become increasingly challenging, posing significant obstacles to the fields of information security and digital forensics. Although existing deep learning methods have achieved certain progress in steganography detection, they still encounter several difficulties in real-world applications. Specifically, current methods often struggle to accurately focus on steganography sensitive regions, leading to limited detection accuracy. Moreover, feature information is frequently lost during transmission, which further reduces the model’s generalization ability. These issues not only compromise the reliability of steganography detection but also hinder its applicability in complex scenarios. To address these challenges, this paper proposes a novel deep image steganalysis network designed to enhance detection accuracy and improve the retention of steganographic information through multilevel feature optimization and global perceptual modeling. The network consists of three core modules: the preprocessing module, the feature extraction module, and the classification module. In the preprocessing stage, a Spatial Rich Model (SRM) filter is introduced to extract the high-frequency residual information of the image to initially enhance the steganographic features; at the same time, a lightweight Densely Connected Convolutional Networks (DenseNet) structure is proposed to enhance the effective transmission and retention of the features and alleviate the information loss problem in the deep network. In the feature extraction stage, a hybrid modeling structure combining depth-separated convolution and ordinary convolution is constructed to improve the feature extraction efficiency and feature description capability; in addition, a dual-domain adaptive attention mechanism integrating channel and spatial dimensions is designed to dynamically allocate feature weights to achieve precise focusing on the steganography-sensitive region. Finally, the classification module adopts dual fully connected layers to realize the effective differentiation between coverage and steganography maps. These innovative designs not only effectively improve the accuracy and generalization ability of steganography detection, but also provide a new efficient network structure for the field of steganalysis. Numerous experimental results show that the detection performance of the proposed method outperforms the existing mainstream methods, such as SR-Net, TSNet, and CVTStego-Net, on the publicly available dataset BOSSbase and BOSW2. Meanwhile, multiple ablation experiments further validate the validity and reasonableness of the proposed network structure. These results not only promote the development of steganalysis technology but also provide more reliable detection tools for the fields of information security and digital forensics.

The fast-paced development of blockchain technology is evident. Yet, the security concerns of smart contracts represent a significant challenge to the stability and dependability of the entire blockchain ecosystem. Conventional smart contract vulnerability detection primarily relies on static analysis tools, which are less efficient and accurate. Although deep learning methods have improved detection efficiency, they are unable to fully utilize the static relationships within contracts. Therefore, we have adopted the advantages of the above two methods, combining feature extraction mode of tools with deep learning techniques. Firstly, we have constructed corresponding feature extraction mode for different vulnerabilities, which are used to extract feature graphs from the source code of smart contracts. Then, the node features in feature graphs are fed into a graph convolutional neural network for training, and the edge features are processed using a method that combines attention mechanism with gated units. Ultimately, the revised node features and edge features are concatenated through a multi-head attention mechanism. The result of the splicing is a global representation of the entire feature graph. Our method was tested on three types of data: Timestamp vulnerabilities, reentrancy vulnerabilities, and access control vulnerabilities, where the F1 score of our method reaches 84.63%, 92.55%, and 61.36%. The results indicate that our method surpasses most others in detecting smart contract vulnerabilities.