Page 422 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

To overcome the third drawback, this work builds an AIGC moderation platform and designs
a multi-level alignment mechanism for safety standards, thereby enhancing the regulatory
compliance of generated content. Moreover, by leveraging the powerful capabilities of
multimodal large language models, the platform not only detects content authenticity but
also identifies manipulated regions, generates heatmap-based visualizations and provides
textual explanations of the model's judgment criteria to improve interpretability and user trust.

Proposal

Firstly, based on mainstream generative models such as Midjourney, GLIDE, and Stable
Diffusion, this case will construct a diverse AIGC dataset. Firstly, this work leverages mainstream
generative models to achieve comprehensive multimodal content coverage. For the text
modality, models such as DeepSeek V3 and Llama 4.0 are used to generate diverse types of
content, including news, commentary, and popular science articles. For the image modality,
models like GPT-4o and Imagen 4 are employed to produce images with varying resolutions,
subjects, and artistic styles. For the video modality, generative models such as Veo 3 and
Gen-4 are utilized to synthesize videos across different scenes and artistic styles. For the audio
modality, models like ACE-Step and Stable Audio are used to generate speech with multilingual,
gender-diverse, and emotionally varied characteristics. Secondly, a wide range of prompts is
designed to enhance the diversity and complexity of the generated content, covering various
cultural contexts, neutral and sensitive topics, as well as high-risk and ethically challenging
scenarios. Finally, each sample is accompanied by detailed metadata, including the source
model, input prompt, content type, post-processing operations, and multimodal association
identifiers, ensuring the dataset’s utility for downstream moderation and forensic research.
The dataset includes various content types, styles, and scenarios, providing rich training data
to support subsequent moderation research. Next, this case adopts an unsupervised domain
adaptation strategy, extracting and integrating multimodal features (text, image, audio, etc.)
to study general multimodal forgery detection methods, enhancing both the accuracy and
applicability of detection. Additionally, the case designs a multi-layered safety standard
alignment mechanism to balance global commonality with regional characteristics. The first
layer involves globally universal standard moderation, while the second layer focuses on region-
specific customized standard moderation. Through this hierarchical mechanism, it ensures both
global compliance of AIGC content and flexibility in adapting to the cultural and social needs of
different regions. Finally, based on the forgery detection and content moderation models, an
AIGC moderation platform will be developed. This platform will collaborate with social media
platforms to promote its demonstration application and industrialization. The platform will
continuously update its data and iteratively improve model performance based on test results.
Moreover, it will support customized moderation for different scenarios. Furthermore, the case
will apply for industry standards to increase the adoption and influence of AIGC moderation.
• Content Moderation: Social media platforms can utilize the proposed AIGC moderation
platform to detect and identify misinformation, implement appropriate countermeasures
to limit its spread, and protect users from deception and harm. This helps ensure the
authenticity and credibility of disseminated AIGC content, fostering a safe and trustworthy
online environment for users.
• Figure Protection: The proposed AIGC moderation platform is capable of detecting
forgery videos or audio recordings of political leaders, as well as identifying maliciously
altered images, fabricated news, and misleading statements. By preventing the spread of
such content during major political events such as elections, the platform helps safeguard

386