Deep forgery detection plays a crucial role in addressing the challenges posed by the rapid spread of deeply generated content that significantly erodes public trust in online information and media. Deeply forged images typically present subtle but significant artifacts in multiple regions, such as in the background, lighting, and localized details. These artifacts manifest as unnatural visual distortions, inconsistent lighting, or irregularities in subtle features that break the natural coherence of the real image. To address these features of forged images, we propose a novel and efficient deep image forgery detection method that utilizes Multi-Graph Attention (MGA) techniques to extract global and local features and minimize accuracy loss. Specifically, our method introduces an interactive dual-channel encoder (DIRM), which aims to extract global and channel-specific features and facilitate complex interactions between these feature sets. In the decoding phase, one of the channels is processed as a block and combined with a Dynamic Graph Attention Network (PDGAN), which is capable of recognizing and amplifying forged traces in local information. To further enhance the model’s ability to capture global context, we propose a global Height–Width Graph Attention Module (HWGAN), which effectively extracts and associates global spatial features. Experimental results show that the classification accuracy of our method for forged images in the GenImage and CIFAKE datasets is comparable to that of the optimal benchmark method. Notably, our model achieves 97.89% accuracy on the CIFAKE dataset and has the lowest number of model parameters and lowest computational overhead. These results highlight the potential of our method for deep forgery image detection.

Generative models based on Variational Autoencoders (VAEs) represent an important area of research in Controllable Text Generation (CTG). However, existing approaches often do not fully exploit the potential of latent variables, leading to limitations in both the diversity and thematic consistency of the generated text. To overcome these challenges, this paper introduces a new framework based on Hierarchical Latent Modulation (HLM). The framework incorporates a hierarchical latent space modulation module for the generation and embedding of conditional modulation parameters. By using low-rank tensor factorization (LMF), the approach combines multi-layer latent variables and generates modulation parameters based on conditional labels, enabling precise control over the features during text generation. Additionally, layer-by-layer normalization and random dropout mechanisms are employed to address issues such as the under-utilization of conditional information and the collapse of generative patterns. We performed experiments on five baseline models based on VAEs for conditional generation, and the results demonstrate the effectiveness of the proposed framework.