Abstract Artificial intelligence technology based on deep learning has been widely used in key fields such as automatic driving, medical diagnosis and financial risk control. These applications also bring more and more serious security problems. In particular, as the means of attack continue to evolve, well-designed countermeasures seriously threaten the reliability of the model and the security of the system. In order to deal with this risk, defensive confrontation samples have become the core task of AI security research, playing a key role in improving the security and credibility of the model. Aiming at the problems of unclear concepts and overlapping standards in previous classification methods, this paper proposes a clearer and unified classification framework, combs and defines the contents of existing research, and solves the inconsistencies. The framework systematically divides the existing countermeasures and defense methods into three categories: detection, purification and optimization. This classification will help researchers understand the actual effects of different methods in the face of various attacks more clearly. This paper also analyzes the tradeoffs between accuracy, robustness, operational efficiency and generalization capability of various defense mechanisms, and reveals how they balance the calculation cost and actual deployment requirements. In addition, the paper points out the main challenges facing the current research, and puts forward the future research directions, including developing more efficient, adaptive, and cross modal defense methods to comprehensively improve the security of AI systems. The purpose of this review is to help researchers understand the development process of anti sample defense technology and provide a reference path for building a stable and reliable AI system.

Steganography without embedding (SWE), which does not modify the cover content, can effectively resist steganalysis and is a highly secure covert communication technology suitable for the Internet of Things (IoT). However, despite this fundamental security advantage, its practical implementation over lossy network channels still faces challenges, including hiding capacity limitations imposed by dataset scales and the non-trivial trade-off between robustness and behavioral security. To tackle these challenges, we propose DIGM-SWE, a robust SWE method based on dynamic image generation and matching. First, a text-to-image generation model (T2I model) is employed to dynamically generate a theme image dataset, ensuring controllable hiding capacity and secure transmission behavior. Second, an image matrix is constructed by distance-pixel cascading sorting, and a bijective mapping is established between secret message sequences and cover image groups via row-index encoding to achieve high hiding success and non-repetitive use of cover images. Finally, resolution enhancement technology is applied to maintain the consistency of stego images under lossy network channels. Meanwhile, a VP-tree structure of the image dataset is built to improve both the stego image matching accuracy and the secret message extraction accuracy. Experimental results show that, compared to other SWE methods, our method achieves up to 3 times greater actual hiding capacity while guaranteeing a 100% hiding success rate across four experimental datasets. Against common image processing operations, it improves average extraction accuracy by up to 59.44%. In two real-world transmission scenarios, our proposed DIGM-SWE achieves perfect recovery of the secret message.

The steganography detection method based on deep learning fuse feature extraction and classification into one model, which reduces the human intervention in feature extraction and obtains a higher detection accuracy than the traditional steganography detection method based on manual feature extraction. However, many existing steganography detection methods based on deep learning generally need to increase the depth and width of the model to further improve the detection accuracy, but it also brings the increase of parameters and Flops (FLoating point OPerations) of the model, resulting in a large consumption of computing resources. Therefore, this manuscript proposes a Convolutional Neural Network(CNN) steganography detection method based on ShuffleBottleneck and Attention mechanism (referred to as ShuffleBANet method, ShuffleBottleneck-Attention-Network Based Method). First, to enhance the network's recognition ability for steganography signals, high-pass filters are used to enhance the steganography feature signals and combined with the improved ShuffleBottleneck structure to enrich the residual features and improve the expression ability of features. Then, a large convolution kernel is used to increase the convolution field of view, and the channel attention mechanism SE (Squeeze and Excitation) and the spatial attention mechanism CA(CoordAttention) are used to capture the residual features between channels and the location information of steganography signals, respectively. Combined with the siamese network framework, the feature extraction backbone is constructed. Finally, the features of the backbone are fused to increase the diversity of features, cover features and stego features are classified using the Softmax function. In this manuscript, BossBase-1.01, BOWS2 and ALASKA#2 datasets are used as cover images, the classical and latest steganalysis methods are used to conduct extensive experiments on the stego images generated by both spatial domain and JPEG domain adaptive steganography algorithms. The experimental results show that compared with the latest SiaStegNet method and the classical SRNet method, the detection efficiency of the proposed method is significantly improved, and the number of parameters and Flops has been reduced by 7.04% / 76.23% and 70.92% / 84.49%, respectively, which provides a solution for the lightweight steganography detection model based on deep learning.

Recent years have witnessed a spurt progress in steganography, which poses challenges for steganalysis. However, previous steganalysis methods attach equal attention to various feature information, while key feature information in detection is ubiquitously ignored, and the detection time-space cost is burdened consequently. To alleviate this predicament, this paper proposes an enhanced image steganalysis via opposition-based evolutionary algorithm (EIS-OBEA), which can guide steganalysis showing more solicitude for key feature information and reduce detection time overhead. Specifically, evolutionary algorithm is introduced into enhanced steganalysis. To elevate searching ability for steganalysis key feature submodels, Tent map is applied in enhanced steganalysis population initialization because of its great randomness. Secondly, considering that opposition-based learning can dynamically adjust searching space of enhanced steganalysis population, opposition-based learning via lens imaging strategy is proposed to help enhanced steganalysis escape from local optimal solutions. Then, to reasonably evaluate the detection contribution of steganalysis key feature submodels, the pearson correlation coefficient for steganalysis is designed. On this basis, fitness function is devised to select superior individuals and obtain steganalysis key feature submodels after iteration. It is noted that EIS-OBEA can optimize steganalysis training samples into quite small-size data, so that computational cost can be significantly reduced when maintaining or even improving detection accuracy. Extensive experimental results substantiate that compared with the state-of-the-art peer algorithms, EIS-OBEA not only achieves highly competitive or even better detection performance, but also meliorates steganalysis time-space cost to a large extent.

The flourishing online social networks provide natural and ideal channels for covert communication, especially image batch steganography, which is characterized by high capacity and efficiency. To address the challenges of covert and reliable messaging, an image robust batch steganography framework (Multi-Stega) is proposed. Utilizing separable steganalysis feature selection and difference measurement, Multi-Stega first designs an embedding sign function to describe steganographic distortion, aiming directly at improving resistance against steganalysis. Then, Multi-Stega applies a cover selection algorithm based on steganographic fitness and a payload distribution strategy based on multi-stage decision optimization, to realize message allocation with minimum embedding signs. On this basis, Multi-Stega can employ any image robust steganography algorithm and universal joint source-channel code to facilitate message embedding and extraction. To analyze its validity, instances are implemented and compared with some state-of-the-art algorithms. Experimental results demonstrate that the separable feature selection provides strong support for precise embedding signs measurement, and Multi-Stega can enhance the detection resistance of representative robust steganography algorithms by 35.10% on average. Covert communication tests on Facebook and Weibo also indicate the concealment and reliability of Multi-Stega, which shows the prospect of practical applications.

Steganalysis feature selection shows excellent effectiveness on elevating the detection efficiency and decreasing time-space cost. However, the single evaluation criterion for features and the subjective selection basis always lead to valuable features neglect, which restricts the improvement of detection accuracy. To alleviate this predicament, this paper proposes a steganalysis feature selection method based on multidimensional evaluation and dynamic threshold allocation (MEDTA method). Firstly, to measure the feature components’ contribution degree to detection, the concept of partial entropy for steganalysis (ste-$pe$) is defined and utilized to measure the mutual information between feature components. On this basis, the evaluation criterion for steganalysis feature components’ contribution degree is proposed, and the theoretical basis is given. Secondly, to measure the functional similarity of the feature components in distinguishing between cover images and stego images, by applying the property of cosine similarity between vectors, the evaluation criterion for steganalysis feature components’ contribution angle is proposed. Then, according to the Occam’s Razor, a multidimensional evaluation criterion based on contribution degree and contribution angle is proposed, which provides a basis for feature selection. In addition, to allocate the threshold for feature selection, this paper proposed a dynamic threshold allocation model, which combines the merits of several function models. Finally, feature selection with multidimensional evaluation and dynamic threshold allocation is proposed, which can achieve a comprehensive evaluation and objective selection for steganalysis features. Extensive experiments conducted on the BOSSbase1.01 image database demonstrate that the proposed MEDTA method could not only achieve highly competitive or even better performance in detection accuracy and feature dimension reduction, as compared with the state-of-the-art methods, but also get rid of depending on classifiers, so that the efficiency of feature extraction and steganalysis gets promoted.

The covert communication technology represented by image steganography can realize the safe and reliable information transmission over an open network. Steganalysis is an important way to measure the security of steganography. The application of deep learning in the field of steganalysis has begun to emerge and achieved certain results. However, the existing steganalysis methods based on deep learning extract limited global features and the detection accuracy still needs to be improved. Therefore, this manuscript proposes an improving CoatNet for spatial and JPEG domain steganalysis (IMCoatNet). Firstly, the SKAttention structure is used to extract fine-grained features. Then, the convolution layer, Transformer layer and the pyramid squeezed attention is used to capture multi-scale spatial features. Finally, a weighted loss is designed to enhance the expression ability of the model. The results show that compared with the traditional Spatial Rich Model (SRM) method, the classical deep residual network for steganalysis (SRNet), and the latest siamese CNN for steganalysis (SiaStegNet), the detection accuracy of the proposed CoatNet method for WOW steganography is improved by 19.61%, 4.33% and 5.04%, respectively; The proposed IMCoatNet is also superior on JPEG domain detection.

Abstract The steganalysis for JPEG image is an important research topic, as the enormous popularity of JPEG image on Internet. However, the stego noise feature extraction process of the existing deep learning‐based steganalytic methods are not adaptive enough to the content of the image, which may lead to suboptimal steganalysis performance. In order to solve this issue, an adaptive stego noise extraction network, named SNENet, for JPEG image steganalysis is proposed. The stego noise extraction module of the network is specifically designed for steganalysis, which consists of parallel dilated convolutional layer and inverted bottleneck layer. This specific design expands the receptive field of the network, which makes the extraction of the stego noise more global and adaptive to the content of the image. The experimental results indicate that proposed network outperforms the state‐of‐the‐art steganalytic method by as much as 6.25% for UED‐JC and 3.35% for J‐UNIWARD. The design of the network is also justified in the extensive ablation experiments.

As the ultimate goal of steganalysis, secret message extraction plays a decisive role in obtaining secret communication evidence and cracking down on criminal activities. For STC (Syndrome-Trellis Codes)-based adaptive steganography, existing pioneering work on secret message extraction: the method based on run test under plaintext embedding may misjudge incorrect stego key as correct stego key, resulting in the failure of extraction. To avoid such a situation, this manuscript proposed a secret message extraction method based on optimal hypothesis test with 100% accuracy under plaintext embedding. First, it is proved that there is a probability distribution difference between the sub-sequence extracted by correct and incorrect stego key. Then, based on the difference, an optimal hypothesis test model is designed to recover the correct stego key. Finally, given the probability of type I and II errors, the sample size and threshold in the hypothesis test are derived. Classic adaptive steganography such as HUGO (Highly Undetectable Steganography) and J-UNIWARD (JPEG Universal Wavelet Relative Distortion) have been conducted experiment, showing that the proposed method can extract message with 100% accuracy and 44 bits sample size, which verifies the correctness of the theorem and the effectiveness of the method.

With the rapid proliferation of urbanization, massive data in social networks are collected and aggregated in real time, making it possible for criminals to use images as a cover to spread secret information on the Internet. How to determine whether these images contain secret information is a huge challenge for multimedia computing security. The steganalysis method based on deep learning can effectively judge whether the pictures transmitted on the Internet in urban scenes contain secret information, which is of great significance to safeguarding national and social security. Image steganalysis based on deep learning has powerful learning ability and classification ability, and its detection accuracy of steganography images has surpassed that of traditional steganalysis based on manual feature extraction. In recent years, it has become a hot topic of the information hiding technology. However, the detection accuracy of existing deep learning based steganalysis methods still needs to be improved, especially when detecting arbitrary-size and multi-source images, their detection efficientness is easily affected by cover mismatch. In this manuscript, we propose a steganalysis method based on Inverse Residuals structured Siamese network (abbreviated as SiaIRNet method, Sia mese- I nverted- R esiduals- Net work Based method). The SiaIRNet method uses a siamese convolutional neural network (CNN) to obtain the residual features of subgraphs, including three stages of preprocessing, feature extraction, and classification. Firstly, a preprocessing layer with high-pass filters combined with depth-wise separable convolution is designed to more accurately capture the correlation of residuals between feature channels, which can help capture rich and effective residual features. Then, a feature extraction layer based on the Inverse Residuals structure is proposed, which improves the ability of the model to obtain residual features by expanding channels and reusing features. Finally, a fully connected layer is used to classify the cover image and the stego image features. Utilizing three general datasets, BossBase-1.01, BOWS2, and ALASKA#2, as cover images, a large number of experiments are conducted comparing with the state-of-the-art steganalysis methods. The experimental results show that compared with the classical SID method and the latest SiaStegNet method, the detection accuracy of the proposed method for 15 arbitrary-size images is improved by 15.96% and 5.86% on average, respectively, which verifies the higher detection accuracy and better adaptability of the proposed method to multi-source and arbitrary-size images in urban scenes.

Traditional steganography methods are usually designed on a lossless channel; thus, messages are often not extracted correctly from an image transmitted over a lossy channel that includes attacks such as scaling. To address this issue, in recent years, the field of robust steganography has emerged. In this paper, the process of image scaling by interpolation is first observed and serves as the basis for proposing the idea of inverse interpolation. Subsequently, the idea of constructing an inverse interpolation equation set is proposed to solve the problem of intersectional blocks during the inverse interpolation process. Then, the scaling factor’s valid range of inverse interpolation is analyzed. Next, the inverse interpolation is successfully applied in robust image steganography. A method that combines antiscaling and antidetection is proposed. Afterward, actual tests on the top 9 mobile phone brands with 28 models and 2 social communication apps that are currently popular in China are done. The scaling factor’s valid range of the proposed method is verified to match the actual lossy channel. The experimental results show that the proposed method achieves a reliable extraction of embedded messages for common interpolation scaling attacks while maintaining high statistical detection resistance.

Cloud services have become an increasingly popular solution to provide different services to clients. More and more data are outsourced to the cloud for storage and computing. With this comes concern about the security of outsourced data. In recent years, homomorphic encryption, blockchain, steganography, and other technologies have been applied to the security and forensics of outsourced data. While encryption technologies such as homomorphic encryption and blockchain scramble data so that they cannot be understood, steganography hides the data so that they cannot be observed. Traditional steganography assumes that the environment is lossless. Robust steganography is grounded in traditional steganography and is proposed based on a real lossy social network environment. Thus, researchers, who study robust steganography, believe that the measurement should follow traditional steganography. However, the application scenario of robust steganography breaks through the traditional default lossless environment premise. It brings about changes in the focus of steganography algorithms. Simultaneously, the existing steganography methods miss the evaluation of applicability and ease of use. In this paper, “default parameters” are observed by comparing the process of robust image steganography with traditional image steganography. The idea of “perfecting default parameters” is proposed. Based on this, the attribute set of measuring robust image steganography is presented. We call it PRUDA (Payload, Robustness, ease of Use, antiDetection, and Applicability). PRUDA perfects default parameters observed in the process of traditional steganography algorithms. Statistics on image processing attacks in mobile social apps and analyses on existing algorithms have verified that PRUDA is reasonable and can better measure a robust steganography method in practical application scenarios.

Considering that traditional image steganography technologies suffer from the potential risk of failure under lossy channels, an enhanced adaptive steganography with multiple robustness against image processing attacks is proposed, while maintaining good detection resistance. First, a robust domain constructing method is proposed utilizing robust element extraction and optimal element modification, which can be applied to both spatial and JPEG images. Then, a robust steganography is proposed based on “Robust Domain Constructing + RS-STC Codes,” combined with cover selection, robust cover extraction, message coding, and embedding with minimized costs. In addition, to provide a theoretical basis for message extraction integrity, the fault tolerance of the proposed algorithm is deduced using error model based on burst errors and decoding damage. Finally, on the basis of parameter discussion about robust domain construction, performance experiments are conducted, and the recommended coding parameters are given for lossy channels with different attacks using the analytic results for fault tolerance. A series of experimental results demonstrate that the proposed algorithm can extract embedded messages with significantly higher accuracy after different attacks, such as compression, noising, scaling and other attacks, compared with the state-of-the-art adaptive steganography, and robust watermarking algorithms, while maintaining good detection resistant performance.

Semantic labeling in high-resolution aerial images is important for its wide range of applications. In this paper, we propose an end-to-end dual multi-scale feature fusion network (DMFFNet) for high-resolution aerial multi-source images. DMFFNet aims to further improve the semantic labeling results of the region where the multispectral features are indistinguishable. Specifically, we design a channel fusion strengthen (CFS) module, which can fuse features adaptively by modelling interdependencies between channels. Furthermore, a multiscale context aggregation (MCA) module is utilized to obtain larger receptive field and more contextual information. The experiment results confirm the DMFFNet with CFS and MCA improve the semantic labeling performance by utilizing multi-source data.

The performance of object instance segmentation in remote sensing images has been greatly improved through the introduction of many landmark frameworks based on convolutional neural network. However, the object densely issue still affects the accuracy of such segmentation frameworks. Objects of the same class are easily confused, which is most likely due to the close docking between objects. We think context information is critical to address this issue. So, we propose a novel framework called SLCMASK-Net, in which a sequence local context module (SLC) is introduced to avoid confusion between objects of the same class. The SLC module applies a sequence of dilation convolution blocks to progressively learn multi-scale context information in the mask branch. Besides, we try to add SLC module to different locations in our framework and experiment with the effect of different parameter settings. Comparative experiments are conducted on remote sensing images acquired by QuickBird with a resolution of 0.5m - 1m and the results show that the proposed method achieves state-of-the-art performance.

Steganography detection based on Rich Model features is a hot research direction in steganalysis. However, rich model features usually result a large computation cost. To reduce the dimension of steganalysis features and improve the efficiency of steganalysis algorithm, differing from previous works that normally proposed new feature extraction algorithm, this paper proposes a general steganalysis feature selection method based on decision rough set α-positive region reduction. First, it is pointed out that decision rough set α-positive region reduction is suitable for steganalysis feature selection. Second, a quantization method of attribute separability is proposed to measure the separability of steganalysis feature components. Third, steganalysis feature components selection algorithm based on decision rough set α-positive region reduction is given; thus, stego images can be detected by the selected feature. The proposed method can significantly reduce the feature dimensions and maintain detection accuracy. Based on the BOSSbase-1.01 image database of 10000 images, a series of feature selection experiments are carried on two kinds of typical rich model features (35263-D J+SRM feature and 17000-D GFR feature). The results show that even though these two kinds of features are reduced to approximately 8000-D, the detection performance of steganalysis algorithms based on the selected features are also maintained with that of original features, which will remarkably improve the efficiency of feature extraction and stego image detection.

Cloud and cloud shadow detection in remote sensing imagery is important for its wide range of applications. Traditionally, the detection is usually based on the manually designed thresholds from multiband, which is complicated and of multistage. To simplify the process of cloud and cloud shadow detection and improve the performance, we propose a multilevel feature fused segmentation network (MFFSNet), which can be trained end-to-end without any hand-tuned parameters. Specifically, a fully convolutional network is proposed for cloud and cloud shadow features learning. Then, we utilize a novel pyramid pooling module to extract contextual relation between cloud and shadow. Furthermore, a special multilevel feature fused structure is designed to combine semantic information with spatial information from different levels, so that we can better handle the multiscale objects and produce detailed segmentation boundaries. Experiments show that the MFFSNet outperforms the state-of-the-art methods and achieves high accuracies of 98.69% and 98.92% for cloud and cloud shadow detection.

With the dimension of steganalysis feature increases rapidly, ensemble steganalysis has become the trend, and its performance is greatly influenced by the selection of feature subspaces. In order to select feature subspaces more effectively to improve the performance of ensemble steganalysis, a feature subspace selection algorithm based on Fisher criterion is proposed. The proposed selection algorithm computes weight for each feature component according to its Fisher criterion value and a base probability value, then selects the feature components with the probabilities in proportion to their weights. When it is used to improve the ensemble steganalysis, the appropriate base probability value is searched by steps. Experimental results show that for J-UNIWARD (JPEG UNIversal WAvelet Relative Distortion) steganography, the proposed feature subspace selection algorithm can select more effective feature subspaces, and enhance the detection performance of GFR (Gabor Filter Residual) feature.

The CAPTCHA has become an important issue in multimedia security. Aimed at a commonly used text-based CAPTCHA, this paper outlines some typical methods and summarizes the technological progress in text-based CAPTCHA breaking. First, the paper presents a comprehensive review of recent developments in the text-based CAPTCHA breaking field. Second, a framework of text-based CAPTCHA breaking technique is proposed. And the framework mainly consists of preprocessing, segmentation, combination, recognition, postprocessing, and other modules. Third, the research progress of the technique involved in each module is introduced, and some typical methods of segmentation and recognition are compared and analyzed. Lastly, the paper discusses some problems worth further research.

Abstract Current typical adaptive steganography algorithms take the detection resistant capability into account adequately but usually cannot extract the embedded secret messages correctly when stego images suffer from compression attack. In order to solve this problem, a framework of adaptive steganography resisting JPEG compression and detection is proposed. Utilizing the relationship between Discrete Cosine Transformation (DCT) coefficients, the domain of messages embedding is determined; for the maximum of the JPEG compression resistant ability, the modifying magnitude of different DCT coefficients caused by messages embedding can be determined; in order to ensure the completely correct extraction of embedded messages after JPEG compression, error correct codes are used to encode the messages to be embedded; on the basis of the current distortion functions, the distortion value of DCT coefficients corresponding to the modifying magnitude in the embedding domain can be calculated; to improve the detection resistant ability of the stego images and realize the minimum distortion embedding, syndrome‐trellis codes are used to embed the encoded messages into the DCT coefficients that have a smaller distortion value. Based on the proposed framework, an adaptive steganography algorithm resisting JPEG compression and detection is designed, which utilizes the relationship between coefficients in a DCT block and the means of that in three adjacent DCT blocks. The experimental results that demonstrate the proposed algorithm not only has a good JPEG compression resistant ability but also has a strong detection resistant performance. Comparing with current J‐UNIWARD steganography under quality factor 85 of JPEG compression, the extraction error rates without pre‐compression decrease from about 50% to nearly 0, while the stego images remain a good detection resistant ability comparing with a typical robust watermarking algorithm, which shows the validity of the proposed framework. Copyright © 2016 John Wiley & Sons, Ltd.

Current typical adaptive Steganography algorithms cannot extract the embedded secret messages correctly after compression. In order to solve this problem, a JPEG-compression resistant adaptive steganography algorithm is proposed. Utilizing the relationship between DCT coefficients, the domain of messages embedding is determined. The modifying magnitude of different DCT coefficients can be determined according to the quality factors of JPEG compression. To ensure the completely correct extraction of embedded messages after JPEG compression, the RS codes is used to encode the messages to be embedded. Besides, based on the current energy function in the PQe steganography and the distortion function in J-UNIWARD Steganography, the corresponding distortion value of DCT coefficients is calculated. With the help of that, STCs is used to embed the encoded messages into the DCT coefficients, which have a smaller distortion value. The experimental results under different quality factors of JPEG compression and different payloads demonstrate that the proposed algorithm not only has a high correct rate of extracted messages after JPEG compression, which increases from about 60% to nearly 100% comparing with J-UNIWARD steganography under quality factor 75 of JPEG compression, but also has a strong detection resistant performance.

Adaptive JPEG steganographic schemes are difficult to preserve the image texture features in all scales and orientations when the embedding changes are constrained to the complicated texture regions, then a steganalysis feature extraction method is proposed based on 2 dimensional (2D) Gabor filters. The 2D Gabor filters have certain optimal joint localization properties in the spatial domain and in the spatial frequency domain. They can describe the image texture features from different scales and orientations, therefore the changes of image statistical characteristics caused by steganography embedding can be captured more effectively. For the proposed feature extraction method, the decompressed JPEG image is filtered by 2D Gabor filters with different scales and orientations firstly. Then, the histogram features are extracted from all the filtered images.Lastly, the ensemble classifier is used to assemble the proposed steganalysis feature as well as the final steganalyzer. The experimental results show that the proposed steganalysis feature can achieve a competitive performance by comparing with the other steganalysis features when they are used for the detection performance of adaptive JPEG steganography such as UED, JUNIWARD and SI-UNIWARD.