Jia Liu

Moyan Li, Jinmiao Fu, Shaoyuan Xu et al.

On shopping websites, product images of low quality negatively affect customer experience. Although there are plenty of work in detecting images with different defects, few efforts have been dedicated to correct those defects at scale. A major challenge is that there are thousands of product types and each has specific defects, therefore building defect specific models is unscalable. In this paper, we propose a unified Image-to-Image (I2I) translation model to correct multiple defects across different product types. Our model leverages an attention mechanism to hierarchically incorporate high-level defect groups and specific defect types to guide the network to focus on defect-related image regions. Evaluated on eight public datasets, our model reduces the Frechet Inception Distance (FID) by 24.6% in average compared with MoNCE, the state-of-the-art I2I method. Unlike public data, another practical challenge on shopping websites is that some paired images are of low quality. Therefore we design our model to be semi-paired by combining the L1 loss of paired data with the cycle loss of unpaired data. Tested on a shopping website dataset to correct three image defects, our model reduces (FID) by 63.2% in average compared with WS-I2I, the state-of-the art semi-paired I2I method.

Junkai Chen, Zhijie Deng, Kening Zheng et al.

As Multimodal Large Language Models (MLLMs) develop, their potential security issues have become increasingly prominent. Machine Unlearning (MU), as an effective strategy for forgetting specific knowledge in training data, has been widely used in privacy protection. However, MU for safety in MLLM has yet to be fully explored. To address this issue, we propose SAFEERASER, a safety unlearning benchmark for MLLMs, consisting of 3,000 images and 28.8K VQA pairs. We comprehensively evaluate unlearning methods from two perspectives: forget quality and model utility. Our findings show that existing MU methods struggle to maintain model performance while implementing the forget operation and often suffer from over-forgetting. Hence, we introduce Prompt Decouple (PD) Loss to alleviate over-forgetting through decouple prompt during unlearning process. To quantitatively measure over-forgetting mitigated by PD Loss, we propose a new metric called Safe Answer Refusal Rate (SARR). Experimental results demonstrate that combining PD Loss with existing unlearning methods can effectively prevent over-forgetting and achieve a decrease of 79.5% in the SARR metric of LLaVA-7B and LLaVA-13B, while maintaining forget quality and model utility. Our code and dataset will be released upon acceptance. Warning: This paper contains examples of harmful language and images, and reader discretion is recommended.

Haisong Liu, Tao Lu, Yihui Xu et al.

In this paper, we study the problem of jointly estimating the optical flow and scene flow from synchronized 2D and 3D data. Previous methods either employ a complex pipeline that splits the joint task into independent stages, or fuse 2D and 3D information in an "early-fusion" or "late-fusion" manner. Such one-size-fits-all approaches suffer from a dilemma of failing to fully utilize the characteristic of each modality or to maximize the inter-modality complementarity. To address the problem, we propose a novel end-to-end framework, called CamLiFlow. It consists of 2D and 3D branches with multiple bidirectional connections between them in specific layers. Different from previous work, we apply a point-based 3D branch to better extract the geometric features and design a symmetric learnable operator to fuse dense image features and sparse point features. Experiments show that CamLiFlow achieves better performance with fewer parameters. Our method ranks 1st on the KITTI Scene Flow benchmark, outperforming the previous art with 1/7 parameters. Code is available at https://github.com/MCG-NJU/CamLiFlow.

Junzhuo Li, Bo Wang, Xiuze Zhou et al.

The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored. Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures. To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B). Results show MoE models achieve 37% higher per-layer efficiency via a "mid-activation, late-amplification" pattern: early layers screen experts, while late layers refine knowledge collaboratively. Ablation studies reveal a "basic-refinement" framework--shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes). Semantic-driven routing is evidenced by strong correlations between attention heads and experts (r=0.68), enabling task-aware coordination. Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop). Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation. These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

Jia Liu, Yan Ke, Yu Lei et al.

In the past few years, the Generative Adversarial Network (GAN) which proposed in 2014 has achieved great success. GAN has achieved many research results in the field of computer vision and natural language processing. Image steganography is dedicated to hiding secret messages in digital images, and has achieved the purpose of covert communication. Recently, research on image steganography has demonstrated great potential for using GAN and neural networks. In this paper we review different strategies for steganography such as cover modification, cover selection and cover synthesis by GANs, and discuss the characteristics of these methods as well as evaluation metrics and provide some possible future research directions in image steganography.

Zhuo Zhang, Guangyuan Fu, Fuqiang Di et al.

The traditional reversible data hiding technique is based on cover image modification which inevitably leaves some traces of rewriting that can be more easily analyzed and attacked by the warder. Inspired by the cover synthesis steganography based generative adversarial networks, in this paper, a novel generative reversible data hiding scheme (GRDH) by image translation is proposed. First, an image generator is used to obtain a realistic image, which is used as an input to the image-to-image translation model with CycleGAN. After image translation, a stego image with different semantic information will be obtained. The secret message and the original input image can be recovered separately by a well-trained message extractor and the inverse transform of the image translation. Experimental results have verified the effectiveness of the scheme.

Yan Ke, Min-qing Zhang, Jia Liu et al.

This paper proposes a fully homomorphic encryption encapsulated difference expansion (FHEE-DE) scheme for reversible data hiding in encrypted domain (RDH-ED). In the proposed scheme, we use key-switching and bootstrapping techniques to control the ciphertext extension and decryption failure. To realize the data extraction directly from the encrypted domain without the private key, a key-switching based least-significant-bit (KS-LSB) data hiding method has been designed. In application, the user first encrypts the plaintext and uploads ciphertext to the server. Then the server performs data hiding by FHEE-DE and KS-LSB to obtain the marked ciphertext. Additional data can be extracted directly from the marked ciphertext by the server without the private key. The user can decrypt the marked ciphertext to obtain the marked plaintext. Then additional data or plaintext can be obtained from the marked plaintext by using the standard DE extraction or recovery. A fidelity constraint of DE is introduced to reduce the distortion of the marked plaintext. FHEE-DE enables the server to implement FHEE-DE recovery or extraction on the marked ciphertext, which returns the ciphertext of original plaintext or additional data to the user. In addition, we simplified the homomorphic operations of the proposed universal FHEE-DE to obtain an efficient version. The Experimental results demonstrate that the embedding capacity, fidelity, and reversibility of the proposed scheme are superior to existing RDH-ED methods, and fully separability is achieved without reducing the security of encryption.

Zhuo Zhang, Jia Liu, Yan Ke et al.

In this paper, a novel data-driven information hiding scheme called generative steganography by sampling (GSS) is proposed. Unlike in traditional modification-based steganography, in our method the stego image is directly sampled by a powerful generator: no explicit cover is used. Both parties share a secret key used for message embedding and extraction. The Jensen-Shannon divergence is introduced as a new criterion for evaluating the security of generative steganography. Based on these principles, we propose a simple practical generative steganography method that uses semantic image inpainting. The message is written in advance to an uncorrupted region that needs to be retained in the corrupted image. Then, the corrupted image with the secret message is fed into a Generator trained by a generative adversarial network (GAN) for semantic completion. Message loss and prior loss terms are proposed for penalizing message extraction error and unrealistic stego image. In our design, we first train a generator whose training target is the generation of new data samples from the same distribution as that of existing training data. Next, for the trained generator, backpropagation to the message and prior loss are introduced to optimize the coding of the input noise data for the generator. The presented experiments demonstrate the potential of the proposed framework based on both qualitative and quantitative evaluations of the generated stego images.

Jia Liu, Yan Ke, Yu Lei et al.

In order to keep the data secret, various techniques have been implemented to encrypt and decrypt the secret data. Cryptography is committed to the security of content, i.e. it cannot be restored with a given ciphertext. Steganography is to hiding the existence of a communication channel within a stego. However, it has been difficult to construct a cipher (cypher) that simultaneously satisfy both channel and content security for secure communication. Inspired by the Cardan grille, this paper presents a new generative framework for grille cipher. A digital cardan grille is used for message encryption and decryption. The ciphertext is directly sampled by a powerful generator without an explicit cover. Message loss and prior loss are proposed for penalizing message extraction error and unrealistic ciphertext. Jensen-Shannon Divergence is introduced as new criteria for channel security. A simple practical data-driven grille cipher is proposed using semantic image inpainting and generative adversarial network. Experimental results demonstrate the promising of the proposed method.

Jia Liu, Tanping Zhou, Zhuo Zhang et al.

In this paper, a new framework for construction of Cardan grille for information hiding is proposed. Based on the semantic image inpainting technique, the stego image are driven by secret messages directly. A mask called Digital Cardan Grille (DCG) for determining the hidden location is introduced to hide the message. The message is written to the corrupted region that needs to be filled in the corrupted image in advance. Then the corrupted image with secret message is feeded into a Generative Adversarial Network (GAN) for semantic completion. The adversarial game not only reconstruct the corrupted image , but also generate a stego image which contains the logic rationality of image content. The experimental results verify the feasibility of the proposed method.

Ming-ming Liu, Min-qing Zhang, Jia Liu et al.

Traditional image steganography modifies the content of the image more or less, it is hard to resist the detection of image steganalysis tools. To address this problem, a novel method named generative coverless information hiding method based on generative adversarial networks is proposed in this paper. The main idea of the method is that the class label of generative adversarial networks is replaced with the secret information as a driver to generate hidden image directly, and then extract the secret information from the hidden image through the discriminator. It's the first time that the coverless information hiding is achieved by generative adversarial networks. Compared with the traditional image steganography, this method does not modify the content of the original image. therefore, this method can resist image steganalysis tools effectively. In terms of steganographic capacity, anti-steganalysis, safety and reliability, the experimen shows that this hidden algorithm performs well.

Yan Ke, Minqing Zhang, Jia Liu et al.

The distortion in steganography that usually comes from the modification or recoding on the cover image during the embedding process leaves the steganalyzer with possibility of discriminating. Faced with such a risk, we propose generative steganography with Kerckhoffs' principle (GSK) in this letter. In GSK, the secret messages are generated by a cover image using a generator rather than embedded into the cover, thus resulting in no modifications in the cover. To ensure the security, the generators are trained to meet Kerckhoffs' principle based on generative adversarial networks (GAN). Everything about the GSK system, except the extraction key, is public knowledge for the receivers. The secret messages can be outputted by the generator if and only if the extraction key and the cover image are both inputted. In the generator training procedures, there are two GANs, Message- GAN and Cover-GAN, designed to work jointly making the generated results under the control of the extraction key and the cover image. We provide experimental results on the training process and give an example of the working process by adopting a generator trained on MNIST, which demonstrate that GSK can use a cover image without any modification to generate messages, and without the extraction key or the cover image, only meaningless results would be obtained.