Page 421 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

technology providers to ensure that synthesized audio, video, text, and image content can
be detected as AI-generated. The AI Act also prohibits the use of AI systems deemed to
pose a clear threat to fundamental user rights. Relevant regulations include the European
Commission’s GDPR (2016), China’s Interim Measures for the Management of Generative 4.4-Productivity
Artificial Intelligence Services (2023), and the Global AI Governance Initiative (2023). This work
focuses on moderating AIGC content to ensure compliance with regulations across different
countries and regions, and covers the academic and industrial AIGC products such as GPT-
4o, Imagen 4, Midjourney, Stable Diffusion, Veo 3, Gen-4, Pika, Sora, ACE-Step, Stable Audio,
MuseNet, V2A, DeepSeek, Gemini, Claude, Llama 4.0, etc. The solution will leverage the
advanced artificial intelligence to develop an App, platform, SDK and mini-program to verify
the authenticity and compliance of AIGC content. Existing solutions primarily rely on manual
review or rule-based keyword engines, but these methods are limited by the complexity of
contextual understanding and the ability to process vast amounts of data. For instance,
• Content moderators manually review text, images, and videos to determine whether
the content violates platform policies, such as those related to political sensitivity,
pornography, violence, or misinformation. However, this approach is highly subjective
and costly, making it difficult to scale effectively for large volumes of content.
• The regular expression matching method identifies patterns within input text by defining
a set of rules or patterns using regex syntax. However, it has notable limitations, such as
a lack of semantic understanding and poor readability and maintainability when dealing
with complex regular expressions.
• The DFA (Deterministic Finite Automaton) sensitive word filtering algorithm efficiently
identifies sensitive terms in a given text by matching them against a predefined
keyword database. However, the algorithm has limitations, including a lack of semantic
understanding and an inability to handle complex contextual relationships, which makes
it vulnerable to evasion.

AI-based approaches offer several benefits:

(1) They can handle massive data volumes simultaneously, significantly improving detection
efficiency.
(2) They capture contextual semantic relationships, enhancing detection accuracy.
(3) They integrate multimodal features, further improving detection accuracy.

However, AI-based methods also have drawbacks:
(1) They depend heavily on large datasets with reliable labels, making them vulnerable to
data poisoning attacks and prone to poor generalization.
(2) They lack robustness against post-processing operations like compression and cropping,
which are commonly used on social platforms.
(3) The features extracted by AI models often lack interpretability.

To address the first drawback, this work adopts unsupervised learning to reduce dependence on
large amounts of labeled data and to resist data poisoning attacks. Additionally, by constructing
a diverse AIGC database that covers various types, styles, and scenarios, the generalization
capability of the model is significantly improved.

To tackle the second drawback, this work integrates multimodal features, including audio, visual,
and textual modalities, and applies data augmentation to enhance the model’s robustness
against common post-processing operations such as compression and cropping.

385