Meta AI introduces CLUE (Constitutional MLLM JUdgE) – an AI framework designed to address the shortcomings of traditional image security systems

The rapid growth of digital platforms has highlighted the security of images. Harmful images, ranging from explicit content to depictions of violence, pose significant challenges for content moderation. The proliferation of ai-generated content (AIGC) has exacerbated these challenges, as advanced image generation models can easily create unsafe images. Current security systems rely heavily on human-labeled data sets, which are expensive and difficult to scale. Additionally, these systems often have difficulty adapting to complex and constantly evolving security guidelines. An effective solution must address these limitations while ensuring efficient and reliable image security assessments.

Researchers at Meta, Rutgers University, Westlake University, and UMass Amherst have developed CLUE (Constitutional MLLM JUdgE), a framework designed to address the shortcomings of traditional image security systems. CLUE uses multimodal large language models (MLLM) to convert subjective security rules into objective, measurable criteria. Key features of the framework include:

1. Reification of the Constitution: Convert subjective security rules into clear and actionable guidelines for better processing by MLLMs.
2. Image and rule relevance checks: Leverage CLIP to efficiently filter out irrelevant rules by evaluating relevance between images and guidelines.
3. Extraction of preconditions: Break complex rules into simplified precondition chains to make reasoning easier.
4. Probability analysis of unbiased tokens: Mitigate biases caused by previous languages ​​and non-core image regions to improve objectivity.
5. Cascade reasoning: Employ deeper chain-of-thought reasoning for low-confidence cases to improve decision-making accuracy.

Technical details and benefits

The CLUE framework addresses key challenges associated with MLLMs in image security. By objectifying safety rules, it replaces ambiguous guidelines with precise criteria, such as specifying that “people should not be depicted with visible, bloody wounds indicating imminent death.”

Relevance scanning using CLIP streamlines the process by removing rules irrelevant to the inspected image, reducing the computational burden. This ensures that the framework focuses only on relevant rules, improving efficiency.

The precondition extraction module simplifies complex rules into logical components, allowing MLLMs to reason more effectively. For example, a rule like “must not represent any people whose bodies are on fire” breaks down into conditions like “the people are visible” and “the bodies are on fire.”

The unbiased token probability analysis is another notable feature. By comparing token probabilities with and without image tokens, biases are identified and minimized. This reduces the likelihood of errors, such as associating background elements with violations.

The waterfall reasoning mechanism provides strong support for low-confidence scenarios. Using step-by-step logical reasoning, it ensures accurate assessments, even for edge cases, while providing detailed justifications for decisions.

Experimental results and insights

The effectiveness of CLUE has been validated through extensive testing on several MLLM architectures, including InternVL2-76B, Qwen2-VL-7B-Instruct, and LLaVA-v1.6-34B. Key findings include:

• Precision and recovery: CLUE achieved 95.9% recall and 94.8% precision with InternVL2-76B, outperforming existing methods.
• Efficiency: The relevance scanning module filtered out 67% of irrelevant rules and retained 96.6% of violated rules, significantly improving computational efficiency.
• Generalizability: Unlike refined models, CLUE performed well across various security guidelines, highlighting its scalability.

The ideas also emphasize the importance of the objectification of the constitution and the unbiased analysis of symbolic probability. The objectified rules achieved an accuracy rate of 98.0% compared to 74.0% for their original counterparts, underscoring the value of clear and measurable criteria. Similarly, debiasing improved overall judgment accuracy, with an F1 score of 0.879 for the InternVL2-8B-AWQ model.

Conclusion

CLUE offers a thoughtful and efficient approach to image security, addressing the limitations of traditional methods by leveraging MLLMs. By transforming subjective rules into objective criteria, filtering out irrelevant rules, and using advanced reasoning mechanisms, CLUE provides reliable and scalable solutions for content moderation. Its ability to deliver high accuracy and adaptability makes it a significant advancement in managing the challenges of ai-generated content, paving the way for safer online platforms.

Verify he Paper. All credit for this research goes to the researchers of this project. Also, don't forget to follow us on <a target="_blank" href="https://x.com/intent/follow?screen_name=marktechpost” target=”_blank” rel=”noreferrer noopener”>twitter and join our Telegram channel and LinkedIn Grabove. Don't forget to join our SubReddit over 65,000 ml.

UPCOMING FREE ai WEBINAR (JANUARY 15, 2025): <a target="_blank" href="https://info.gretel.ai/boost-llm-accuracy-with-sd-and-evaluation-intelligence?utm_source=marktechpost&utm_medium=newsletter&utm_campaign=202501_gretel_galileo_webinar”>Increase LLM Accuracy with Synthetic Data and Assessment Intelligence–<a target="_blank" href="https://info.gretel.ai/boost-llm-accuracy-with-sd-and-evaluation-intelligence?utm_source=marktechpost&utm_medium=newsletter&utm_campaign=202501_gretel_galileo_webinar”>Join this webinar to learn practical information to improve LLM model performance and accuracy while protecting data privacy..