The surface texture of objects in industrial scenes is complex and diverse, and the characteristics of surface defects are often very similar to the surrounding environment and texture background, so it is difficult to accurately detect the defect area. However, when deep learning technology is used to detect complex texture surface defects, the detection accuracy is not high, due to the lack of large-scale pixel-level label datasets. Therefore, a defect detection model Siamese-RCNet for complex texture surface with a small number of annotations is proposed. The Cascade R-CNN target detection network is used as the basic framework, making full use of unlabeled image feature information, and fusing the nonlinear relationship learning ability of Siamese network and the feature extraction ability of the Res2Net backbone network to more effectively capture the subtle features of complex texture surface defects. The image difference measurement method is used to calculate the similarity between different images, and the attention module is constructed to weight the feature map of the feature extraction pyramid, so that the model can focus more on the defect area and suppress the influence of complex background texture area, so as to improve the accuracy of detection. To verify the effectiveness of the Siamese-RCNet model, a series of experiments were carried out on the DAGM2007 dataset of weakly supervised learning texture surface defects for industrial optical inspection. The results show that even if only 20% of the labeled datasets are used, the mAP@0.5 of the Siamese-RCNet model can still reach 96.9%. Compared with the traditional Cascade R-CNN and Faster R-CNN target detection networks, the Siamese-RCNet model has high accuracy, can reduce the workload of manual labeling, and provides strong support for practical applications.

Aiming at the fact that the moving target defense game model fails to accurately portray attack and defense gains, resulting in bias in attack and defense games and the inability to select effective defense strategies, we construct the moving target three-way evolutionary game defense model (MTTEGDM). Firstly, the model is defined and analyzed theoretically under the premise of uncertainty and irrationality. Then, combined with the three-way decisions, the attack intention is introduced into the target network loss calculation, and a dynamic weight adjustment algorithm based on the three-way decisions is proposed to accurately characterize the attack and defense gains from a multi-attribute perspective. Finally, the evolutionary game model is used to analyze the evolution trend of the multi-stage defense strategy, so as to carry out feasible and effective defense behavior. The simulation results show that the model can accurately predict the optimal defense strategy of moving targets in different stages. Through a Monte Carlo simulation experiment, the proposed algorithm is compared with the traditional evolutionary game model, and the effectiveness and security of the proposed algorithm are verified.

To address the current issue that existing steganographic techniques are highly dependent on the quality of carrier images and have insufficient adaptability to damaged images, we propose an MAE-based inpainting-steganography framework, which realizes steganography and inpainting for 2D color damaged images through multi-module collaboration. The encoder divides the original 244 × 244 images into non-overlapping patches and performs random masking to generate their feature sequences. The steganographic point prediction module locates suitable steganographic points by analyzing feature fluctuations during the inpainting process. The steganography module embeds information by introducing feature offsets at specific positions in the feature sequences according to the suitable steganographic points. The decoder completes the inpainting of masked regions to obtain stego-inpainted images. And the extraction module recovers the secret information through feature alignment. Experimental results show that with a feature fluctuation amplitude threshold of 0%, an offset of 3%, and an embedding capacity of 30 bits, the verification success rate reaches 96.67%, and remains at 95.27% even when the capacity is increased to 1000 bits. In terms of anti-steganography detection capability, evaluation results based on mainstream steganalysis models show that the detection accuracy of the proposed method is 49.79%. This accuracy is close to the level of random guessing, which indicates the excellent anti-detection performance of the method. Moreover, under the combined perturbations of JPEG compression, Gaussian blur, and noise interference, the information extraction accuracy still reaches 94.68%, demonstrating that the feature-domain embedding mechanism has good robustness. In addition, adding steganography under different masking ratios has no significant impact on image inpainting performance, and some metrics even outperform pure inpainting algorithms. While maintaining the quality of image inpainting, this framework achieves highly reliable information embedding and extraction, and maintains a high verification success rate in the complex scenario of damaged carrier images, providing a new idea for steganographic techniques in practical applications.

Current deep learning models for side-channel analysis are often parameter-inefficient and lack robustness in feature extraction under noisy conditions, thereby requiring a substantial number of power traces to recover the key. We propose Stal-net, a lightweight task-oriented architecture for profiling side-channel analysis. It integrates lightweight local feature extraction, temporal dependency modeling, dynamic-threshold-based feature refinement, and correlation-aware multi-scale fusion in a coordinated pipeline, improving feature discriminability while controlling model complexity. Experiments were conducted on the ASCAD and DPA contest v4 datasets. Key recovery performance is evaluated using classification accuracy and average key rank, and attack efficiency is measured by the average number of power traces required to achieve average key rank of 0. On ASCAD, Stal-net achieves 94.23% accuracy and reaches average key rank = 0 with only 22 power traces, improving accuracy by 1.33% and attack efficiency by 26.7% over the baseline. On DPA contest v4, Stal-net achieves 93.4% accuracy and reaches average key rank = 0 with only 2 power traces, improving accuracy by 5.18% and attack efficiency by 33.3%. Ablation studies further demonstrate the complementary roles of the four core modules in noise suppression, temporal modeling, and feature fusion. This collaborative mechanism is the key enabler for Stal-net to maintain high performance while adhering to a lightweight design.