Desynchronization attacks, such as weight permutations in trained models, hinder the recovery of embedded watermarks and pose a serious threat to white-box watermarking for deep neural networks (DNNs). In this paper, we study white-box watermarking methods that are robust against desynchronization attacks. As naïve approaches, we consider a multiplication-based method and a brute-force method; however, they suffer from two major drawbacks: inaccurate watermark recovery and degradation of model performance. We further propose a multilayered matrix-multiplication (MMM) method that targets feature components defined by matrix products. This design improves robustness against desynchronization attacks while preserving model performance. The MMM method is applicable to a wide range of model architectures. Through simulations, we evaluate these methods from two perspectives: watermark recovery accuracy under desynchronization attacks and the impact of watermark embedding on model performance.

This paper proposes an original approach for the automatic detection of AI-generated images, using features derived from noise residuals artefacts. Contrary to most current research that leverages sophisticated deep learning models to further improve performance, this study highlights the distinct noise residual characteristics in deepfakes, facilitating the identification of AI-generative images. Our findings highlight some limitations of image models, which can be used for forensic analysis and for future AI-based text-to-image generative models. Broad numerical results on a large and diverse dataset show the interest of the identified features as well as the relevance of the present method.

Crack detection, particularly from pavement images, presents a formidable challenge in computer vision due to inherent complexities such as intensity inhomogeneity, intricate topologies, low contrast, and noisy backgrounds. Automated crack detection is crucial for maintaining the structural integrity of essential infrastructures, including buildings, pavements, and bridges. Existing lightweight methods often face challenges, including computational inefficiency, complex crack patterns, and difficult backgrounds, leading to inaccurate detection and impracticality for real-world applications. We propose EfficientCrackNet, a lightweight hybrid model combining Convolutional Neural Networks (CNNs) and transformers for precise crack segmentation to address these limitations. EfficientCrackNet integrates depthwise separable convolutions (DSC) layers and MobileViT block to capture global and local features. The model employs an Edge Extraction Method (EEM) for efficient crack edge detection without pretraining and an Ultra-Lightweight Subspace Attention Module (ULSAM) to enhance feature extraction. Extensive experiments on three benchmark datasets, Crack500, DeepCrack, and GAPs384, demonstrate that EfficientCrackNet achieves superior performance compared to existing lightweight models, requiring only 0.26M parameters and 0.483 GFLOPs. The proposed model offers an optimal balance between accuracy and computational efficiency, outperforming state-of-the-art lightweight models and providing a robust and adaptable solution for real-world crack segmentation.

The widespread use of digital platforms, prioritising content based on engagement metrics and rewarding content creators accordingly, has contributed to the proliferation of disinformation and its far-reaching social and political impact. In addition, digital platforms often operate as black boxes, concealing their decision-making processes from users and prioritizing investor interests over ethical and social considerations. Consequently, this has contributed to the erosion of general trust in verification systems. To mitigate this issue, our project proposes a two-stage verification system. The first stage allows media industries to watermark their image and video content. The second stage involves implementing a machine-learning-based manipulation detection system for suspicious content. We present findings from an international user experience study, where potential online news consumers verified the authenticity of images on a prototype version of our system. In this paper, we reflect on critical issues of explainability addressed by participants in our user study and how we addressed this issue in the platform’s design.

Researchers are increasingly interested to study novel attacks on machine learning models. The classifiers are fooled by making small perturbation to the input or by learning patches that can be applied to objects. In this paper we present an iterative approach to generate a patch that when digitally placed on the face can successfully fool the facial recognition system. We focus on dodging attack where a target face is misidentified as any other face. The proof of concept is show-cased using FGSM and FaceNet face recognition system under the white-box attack. The framework is generic and it can be extended to other noise model and recognition system. It has been evaluated for different - patch size, noise strength, patch location, number of patches and dataset. The experiments shows that the proposed approach can significantly lower the recognition accuracy. Compared to state of the art digital-world attacks, the proposed approach is simpler and can generate inconspicuous natural looking patch with comparable fool rate and smallest patch size.

Digital media have changed the classical model of mass media that considers the transmitter of a message and a passive receiver, to a model where users of the digital media can appropriate the contents, recreate, and circulate them. In this context, online social media are a suitable circuit for the distribution of fake news and the spread of disinformation. Particularly, photo and video editing tools and recent advances in artificial intelligence allow non-professionals to easily counterfeit multimedia documents and create deep fakes. To avoid the spread of disinformation, some online social media deploy methods to filter fake content. Although this can be an effective method, its centralized approach gives an enormous power to the manager of these services. Considering the above, this paper outlines the main principles and research approach of the ongoing DISSIMILAR project, which is focused on the detection of fake news on social media platforms using information hiding techniques, in particular, digital watermarking, combined with machine learning approaches.

For the protection of trained deep neural network(DNN) models, embedding watermarks into the weights of the DNN model have been considered. However, the amount of change in the weights is large in the conventional methods, and it is reported that the existence of hidden watermark can be detected from the analysis of weight variance. This helps attackers to modify the watermark by effectively adding noise to the weight. In this paper, we focus on the fully-connected layers of fine-tuning models and apply a quantization-based watermarking method to the weights sampled from the layers. The advantage of the proposed method is that the change caused by watermark embedding is much smaller and the distortion converges gradually without using any loss function. The validity of the proposed method was evaluated by varying the conditions during the training of DNN model. The results shows the impact of training for DNN model, effectiveness of the embedding method, and high robustness against pruning attacks.

A previously proposed optimal detector for bias-based fingerprinting codes such as Tardos and Nuida requires two kinds of important information: the number of colluders and the collusion strategy used to generate the pirated codeword. An estimator has now been derived for these two parameters. The bias in the pirated codeword is measured by observing the number of zeros and ones and compared with possible bias patterns calculated using information about the collusion strategy and number of colluders. Computer simulation demonstrated that the collusion strategy and number of colluders can be estimated with high probability and that the traceability of a detector using the proposed estimator is extremely close to being optimal.

We consider non-adaptive threshold group testing for identification of up to d defective items in a set of n items, where a test is positive if it contains at least 2 ≤ u ≤ d defective items, and negative otherwise. The defective items can be identified using t = O ((d/u)u(d/d-u)d-u(u log d/u + log 1/∈)·d2log n) tests with probability at least 1 - ∈ for any ∈ > 0 or t = O((d/u)u(d/d-u)d-ud3log n · log d/n) tests with probability 1. The decoding time is t × poly(d2log n). This result significantly improves the best known results for decoding non-adaptive threshold group testing: O(n log n + n log 1/∈) for probabilistic decoding, where ∈ > 0, and O(nulog n) for deterministic decoding.

We consider non-adaptive threshold group testing for identification of up to d defective items in a set of n items, where a test is positive if it contains at least 2 ≤ u ≤ d defective items, and negative otherwise. The defective items can be identified using t=O(( [d/u])u([d/(d-u)])d-u(ulog[d/u]+log[1/(ε)])d2logn) tests with probability at least 1-ε for any or t = O(([b/u])u([d/(d-u)])d-u·d3logn ·log[n/d]) tests with probability 1. The decoding time is t× poly (d2logn). This result significantly improves the best known results for decoding non-adaptive threshold group testing: O(n logn+nlog[1/(ε)]) for probabilistic decoding, where , and O(nulogn) for deterministic decoding.

When a sender side and a receiver side are not synchronized, it is difficult to correctly decode a received codeword. In this study, we investigate a Cyclically Permutable Code (CPC) to immunize the synchronization loss as well as an additive noise over a communication channel. A cyclic code retains the characteristic that cyclically shifted codewords belong to the same code. The codewords in cyclic code can form small cyclic groups such that all codeword in each group is a cyclically shifted version of a certain codeword. The CPC encoder selects each one codeword in each group. Even if a received codeword is cyclically shifted from an original codeword, the receiver can identify the group to which the original codeword belongs. Although there are some methods to generate CPC, the decoding method has not been discussed. Considering the algebraic property of CPC, an efficient decoding method is proposed in this study. The validity of the proposed method is evaluated by simulation.

Robustness against collusion attack has been measured under the marking assumption in fingerprinting codes. However, if codeword symbols are embedded into segments of multimedia content, the assumption must be relaxed. In this letter, we investigate erasures both in segments of multimedia content and in codeword symbols. The former erasure represents the case of a clipping attack, and the latter occurs at the signal-processing domain in which a watermark is embedded. We propose an optimal tracing algorithm under the erasure attack and theoretically estimate the number of erasure symbols in a pirated codeword. To suppress false-positive detections, a cutoff parameter is introduced into the tracing algorithm.

To improve Java programming educations, we have developed a Web-based Java Programming Learning System (JPLAS). To deal with students at different levels, JPLAS provides three levels of problems, namely, element fill-in-blank problems, statement fill-in-blank problems, and code writing problems. Unfortunately, since JPLAS has been implemented by various students who studied in our group at different years, the code has become complex and redundant, which makes further extensions of JPLAS extremely hard. In this paper, we propose the software architecture for JPLAS to avoid redundancy to the utmost at implementations of new functions that will be continued with this JPLAS project. Following the MVC model, our proposal basically uses Java for the model (M), JavaScript/CSS for the view (V), and JSP for the controller (C). For the evaluation, we implement JPLAS by this architecture and compare the number of code files with the previous implementation.

Recently, Java has been educated in many universities and professional schools due to reliability, portability, and scalability. However, because of its limited time in Java programming courses, coding rules are rarely educated. As a result, codes made by students become far from readable codes. In this paper, we propose a coding rule learning function using static code analyzers in Java Programming Learning Assistant System (JPLAS), targeting students who have accomplished the grammar learning and are going to write practical codes for final projects. Coding rules for this function consist of naming rules, coding styles, and potential problems. We evaluate the improvement of readability in four codes refined by this function and the usability of this function through the questionnaires completed by five students. In future studies, we expect that this function will be employed in Java programming courses.

The study of universal detector for fingerprinting code is strongly dependent on the design of scoring function. The best detector is known as the MAP detector that calculates an optimal correlation score, but the number of colluders and their collusion strategy are inevitable. Although there are some scoring functions under some collusion strategies and asymptotic analyses, their numerical evaluation has not been done. In this study, their performance is evaluated for some typical collusion strategies using a discretized bias-based binary fingerprinting code. We also propose a simple but efficient scoring function based on a heuristic observation.

According to watermark security based on the Kerckhoffs' principle, we should mind that illegal users will be able to access to the host signal which is targeted for embedding in a fingerprinting system. In our system, each user's ID is encoded by a fingerprinting code, and then it is embedded into the host signal employing the obfuscation technique. Considering the operation in the obfuscation technique, we discover that the operation is equivalent to a simple spread spectrum (SS) watermarking method using longer sequences. Its effects are intensively evaluated by simulation in this paper.

When a binary fingerprinting codeword is embedded into multimedia content by using a spread-spectrum (SS) watermarking scheme, it is difficult for colluders to perform a symbol-wise attack using their codewords. As discussed in regard to SS fingerprinting schemes, averaging their copies is a cost-effective attack from the signal processing point of view. If the number of colluders is known, an optimal detector can be used against an averaging attack with added white Gaussian noise. The detector first estimates the variance of additive noise, and then calculates correlation scores using a log-likelihood-based approach. However, the number of colluders is not usually known in a real situation. In this paper, we simplify the optimal detector by making statistical approximations and using the characteristics of the parameters for generating codewords. After that, we propose an orthogonal frequency division multiplexing-based SS watermarking scheme to embed the fingerprinting codeword into multimedia content. In a realistic situation, the signal embedded as a fingerprint is in principle attenuated by lossy compression. Because the signal amplitude in a pirated codeword is attenuated, we should adaptively estimate the parameters before calculating the scores. Different from the optimal detector, the simplified detector can easily accommodate changes in signal amplitude by examining the distorted codeword extracted from a pirated copy. We evaluate the performance of the simplified detector through simulation using digital images as well as codewords.

When a binary fingerprinting codeword is embedded into digital contents using a spread-spectrum (SS) watermarking scheme, the marking assumption is not valid anymore because it is difficult for colluders to perform the symbol-wise attack for their codewords. As discussed in the SS-type fingerprinting schemes, veraging their copies is the cost-effective attack from the signal processing point of view. In this paper, we propose an optimal detector under the averaging attack and addition of white Gaussian noise. If the detector knows the number of colluders in advance, it first estimates the variance of additive noise, and then calculates the correlation scores using a log-likelihood-based approach. However, the number of colluders is not given in a real situation. We discover in this study that the characteristic of parameters for generating codewords enables us to eliminate the number of colluders as well as the estimation of the variance of noise at the calculation of correlation score, and propose a simplified detector by analyzing the scoring function in the optimal detector. We evaluate the performance of the simplified detector through simulation using not only codewords, but also a digital image.

The traceability of the spread spectrum fingerprinting has been improved by an iterative detection method combined with an interference removal operation. However, the false-positive probability is slightly increased when the length of fingerprint sequence is rather small. In this study, the iterative detection procedure is adaptively calibrated to maximize the effect of the interference removal operation.

In digital fingerprinting schemes based on the spread spectrum technique, (quasi-)orthogonal sequences are assigned to users as their fingerprints, and they are embedded into digital contents prior to distribution. Owing to the (quasi-) orthogonality, we can uniquely identify the users from a pirated copy, even if dozens of users are involved in the collusion process. Because the number of users accommodated in a fingerprinting system is generally very large, the interference among the sequences involved in a pirated copy becomes non-negligible as the number of colluders increases. In this paper, we investigate the interference from the viewpoint of a communication channel, and we propose an effective removal operation to reduce the interference as far as possible. By iteratively operating the proposed detector, we can sequentially and successively detect colluders from a pirated copy. Furthermore, two kinds of thresholds are introduced in order to perform the removal operation adaptively for detected signals. Experimental results reveal a drastic improvement in the traceability and a reduction in the amount of interference by the removal operation.

Digital fingerprinting is a method to insert user's own ID into digital contents in order to identify illegal users who distribute unauthorized copies.One of the serious problems in a fingerprinting system is the collusion attack such that several users combine their copies of the same content to modify/delete the embedded fingerprints.In this paper, we propose a collusion-resistant fingerprinting scheme based on the CDMA technique.Our fingerprint sequences are orthogonal sequences of DCT basic vectors modulated by PN sequence.In order to increase the number of users, a hierarchical structure is produced by assigning a pair of the fingerprint sequences to a user.Under the assumption that the frequency components of detected sequences modulated by PN sequence follow Gaussian distribution, the design of thresholds and the weighting of parameters are studied to improve the performance.The robustness against collusion attack and the computational costs required for the detection are estimated in our simulation.

Digital fingerprinting of multimedia contents involves the generation of a fingerprint, the embedding operation, and the realization of traceability from redistributed contents.Considering a buyer's right, the asymmetric property in the transaction between a buyer and a seller must be achieved using a cryptographic protocol.In the conventional schemes, the implementation of a watermarking algorithm into the cryptographic protocol is not deeply discussed.In this paper, we propose the method for implementing the spread spectrum watermarking technique in the fingerprinting protocol based on the homomorphic encryption scheme.We first develop a rounding operation which converts real values into integer and its compensation, and then explore the tradeoff between the robustness and communication overhead.Experimental results show that our system can simulate Cox's spread spectrum watermarking method into asymmetric fingerprinting protocol.

In spread spectrum fingerprinting, it has been considered that the strength of the embedded signal is reduced to 1/cof its original value whenccopies are averaged by colluders. In this study, we analyze the model of the averaging attack by considering quantization that causes nonlinear changes in the fingerprint sequence. Our detailed analysis reveals that the attenuation of the signal energy strongly depends on the quantization performed during the embedding and averaging stages. We estimate the actual attenuation factor from the perspective of a stochastic model in the spatial domain and derive an attenuation factor that differs considerably from the conventional one. Our simulation result indicates that the actual attenuation factor is classified into the best and worst cases from the detector's perspective. Furthermore, we demonstrate that colluders can select the worst case by comparing their fingerprinted copies. A countermeasure for preventing the worst-case scenario is also proposed in this paper.

One of the serious problems in a fingerprinting system is the collusion attack such that several users combine their copies of a same content to modify/delete the embedded fingerprints. In IWSEC2007, we have proposed a collusion-resilient fingerprinting scheme based on the CDMA technique. In this paper, we formalize the model of collusion from the viewpoint of a communication channel, and design an effective detector considering the hierarchically embedded signals as a fingerprint.

Digital fingerprinting is used to trace back illegal users, where unique ID known as digital fingerprints is embedded into a content before distribution. On the generation of such fingerprints, one of the important properties is collusion-resistance. Binary codes for fingerprinting with a code length of theoretically minimum order were proposed by Tardos, and the related works mainly focused on the reduction of the code length were presented. In this paper, we present a concrete and systematic construction of the Tardos’s fingerprinting code using a chaotic map. Using a statistical model for correlation scores, a proper threshold for detecting colluders is calculated. Furthermore, for the reduction of computational costs required for the detection, a hierarchical structure is introduced on the codewords. The collusion-resistance of the generated fingerprinting codes is evaluated by a computer simulation.

On the basis of the characteristics of cyclic codes, the codeword space can be partitioned into small subspaces where cyclically shifted codewords of a particular codeword occupy the same subspace. A cyclically permutable code generates codewords belonging to each subspace. However, no approach for the efficient construction of cyclically permutable code from binary cyclic codes has been proposed thus far. In this study, we propose an approach for the efficient and systematic construction of a cyclically permutable code from a cyclic code by utilizing an algebraic property. The proposed coding method improves the robustness of watermarking, particularly for video frames, against a clipping attack

Homomorphic property of public-key cryptosystems is applied for several cryptographic protocols, such as electronic cash, voting system, bidding protocols, etc. Several fingerprinting protocols also exploit the property to achieve an asymmetric system. However, their enciphering rate is extremely low and the implementation of watermarking technique is difficult. In this paper, we propose a new fingerprinting protocol applying additive homomorphic property of Okamoto-Uchiyama encryption scheme. Exploiting the property ingenuously, the enciphering rate of our fingerprinting scheme can be close to the corresponding cryptosystem. We study the problem of implementation of watermarking technique and propose a successful method to embed an encrypted information without knowing the plain value. The security can also be protected for both a buyer and a merchant in our scheme.