The proliferation of multimodal misinformation poses growing threats to public discourse and societal trust. While Large Vision-Language Models (LVLMs) have enabled recent progress in multimodal misinformation detection (MMD), the rise of generative AI (GenAI) tools introduces a new challenge: GenAI-driven news diversity, characterized by highly varied and complex content. We show that this diversity induces multi-level drift, comprising (1) model-level misperception drift, where stylistic variations disrupt a model’s internal reasoning, and (2) evidence-level drift, where expression diversity degrades the quality or relevance of retrieved external evidence. These drifts significantly degrade the robustness of current LVLM-based MMD systems. To systematically study this problem, we introduce DriftBench, a large-scale benchmark comprising 16,000 news instances across six categories of diversification. We design three evaluation tasks: (1) robustness of truth verification under multi-level drift; (2) susceptibility to adversarial evidence contamination generated by GenAI; and (3) analysis of reasoning consistency across diverse inputs. Experiments with six state-of-the-art LVLM-based detectors show substantial performance drops (average F1 ↓ 14.8%) and increasingly unstable reasoning traces, with even more severe failures under adversarial evidence injection. Our findings uncover fundamental vulnerabilities in existing MMD systems and suggest an urgent need for more resilient approaches in the GenAI era.

Dayang Li, Fanxiao Li, Bingbing Song, Li Tang, Wei Zhou. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Abstract Image steganography aims to hide secret data in the cover media for covert communication. Though many deep‐learning‐based image steganography methods have been presented, these approaches suffer from the inefficiency of building long‐distance connections between the cover and secret images, leading to noticeable modification traces and poor steganalysis resistance. To improve the visual imperceptibility of generated stego images, it is essential to establish a global correlation between the cover and secret images. In this way, the secret image can be dispersed throughout the cover image globally. To bridge this gap, a novel image steganography framework called HiiT is proposed, which takes advantage of CNN and Transformer to learn both the local and global pixel correlation in image hiding. Specifically, a new Transformer structure called Inception Transformer is proposed, which incorporates the Inception Net in the attention‐based Transformer architecture. The Inception Net can learn different scaled image features using multiple convolution kernels, while the attention mechanism can learn the global pixel correlation. By this, the proposed Inception Transformer learns the long‐distance pixel dependency between the cover and secret images. Furthermore, we propose a ‘Skip Connection’ mechanism in the proposed Inception Transformer, which merges the low‐level visual features and high‐level semantic features and achieves better model performance. In detail, The HiiT generates higher‐quality stego images with 45.46 PSNR and 0.9915 SSIM. Besides, accurately restored secret images achieve 47.27 PSNR and 0.9952 SSIM. Extensive experimental results show the proposed HiiT significantly improves the image‐hiding performance compared with state‐of‐the‐art methods.

Abstract Image steganography is the art of concealing secret information within images to prevent detection. In deep‐learning‐based image steganography, a common practice is to fuse the secret image with the cover image to directly generate the stego image. However, not all features are equally critical for data hiding, and some insignificant ones may lead to the appearance of residual artifacts in the stego image. In this article, a novel network architecture for image steganography with hybrid attention mechanism based on generative adversarial network is introduced. This model consists of three subnetworks: a generator for generate stego images, an extractor for extracting the secret images, and a discriminator to simulate the detection process, which aids the generator in producing more realistic stego images. A specific hybrid attention mechanism (HAM) module is designed that effectively fuses information across channel and spatial domains, facilitating adaptive feature refinement within deep image representations. The experimental results suggest that the HAM module not only enhances the image quality during both the steganography and extraction processes but also improves the model's undetectability. Stego images are mixed with varying levels of noise in the training process, which can further improve robustness. Finally, it is verified that the model outperforms current steganography approaches on three datasets and exhibits good undetectability.

Most cloud providers offer Deep Learning as a Service (DLaaS) for different business, science and engineering domains. However, it is known that deep neural networks (DNNs) are vulnerable to adversarial examples, which can cause well-trained DNN models to misbehave by injecting human-imperceptible perturbations to the query input data. Securing deep learning as a service becomes a critical challenge in mitigating such adversarial input perturbations, and enhancing the robustness of DNNs. In this article, we report two important facts: First, most adversarial perturbations are high frequency signals or are added to high frequency signals. Second, due to Frequency Principle that neural networks overly pay attention to fit the low frequency signals during training, the models could be easily misled by the high frequency signals of adversarial examples. These facts consequently contribute to the vulnerability of DNNs service in the Cloud. We conjecture that the more robust the neural networks are in learning from high frequency signals, the more resilient these neural networks are against adversarial perturbed examples. We propose a novel method for generating high-frequency-enhanced adversarial examples, which is achieved by a high-pass filter in the frequency domain via Fourier Transform. This method enhances the learning ability for high frequency signals and ameliorates to over-fit useless low frequency signals. In order to improve the robustness of DNNs service under such signal frequency attacks, we propose a multi-modal collaborative adversarial training framework, named as MMCAT, which uses the multi-modal information of the input images for cross-modal collaborative training, delivering excellent extension for effectively learning of multi-modal image information. Extensive experiments show that under strong adaptive frequency attacks, the DNNs service trained with the proposed MMCAT method achieve superior performance and robustness over the state-of-the-art adversarial training approaches.

Recent works on adversarial attacks uncover the intrinsic vulnerability of neural networks, which reveal a critical issue that the neural networks are easily misled by adversarial attacks. As the development of edge computing, more and more real-time tasks are deployed on edge devices. The safety of these neural network-based applications is threatened by adversarial attack. Therefore, the defense technique against adversarial attack has very important application value for edges. Especially, the defense technique should consider the deployment condition on edges, such as low power and low time consumption. Unfortunately, until now, very limited research considers the security problem under adversarial attack on edges. In this paper, we propose a low-power robust learning framework to deal with the adversarial attacks at resource-constrained edge devices. In this framework, we make a rough categorization of approaches on defending against adversarial attacks, and reveal how this edge device-based framework can be used to resist adversarial attacks. Furthermore, we propose a staged ensemble defense strategy in the framework, which achieves better defensive performance than a single defense algorithm. To verify our framework on real application, we build a Drone Search and Rescue System (DSRS) which is employed to examine the performance of the proposed framework. The results indicate that our framework achieves outstanding performance in all aspects, such as robustness, time and power consumption. Multiple evaluations of the low-power robust learning framework provide the advice that help to choose the optimal security configuration on power-constrained and performance-expected environments.