Abstract
This paper presents a novel framework for image safeguarding that addresses the dual needs of transparency and context preservation in content moderation. By integrating a Conditional Vision-Language Model (ConditionalVLM) for grounded rationale generation and a counterfactual reasoning module for surgical obfuscation, the system identifies and masks only the minimum necessary regions of unsafe images. Evaluations on real-world datasets for sexually explicit content, cyberbullying, and self-harm demonstrate that this approach significantly outperforms existing vision-language models in accuracy while maintaining vital context for human review.
Introduction
Content moderation on social media is a high-stakes balancing act. When an image is flagged as unsafe, platforms must provide specific rationales for removal while preserving enough context for human moderators and law enforcement to identify victims or locations. Traditional "blanket blurring" fails on both counts, often hiding evidence like tattoos or geographical landmarks. This research introduces a two-step framework designed to provide precise, attribute-grounded explanations and perform "surgical" obfuscation that hides harm without stripping away essential context.
Methods
The proposed framework consists of two primary modules:
1. ConditionalVLM (The Reasoning Module)
To provide accurate rationales, the researchers developed ConditionalVLM, which features a Conditional Image Instruction-guided Transformer (CIIT). Unlike generic models, ConditionalVLM conditions language generation on the output of pre-trained unsafe image classifiers. These classifiers act as control codes, forcing the model to focus on domain-specific attributes—such as "rude hand gestures"—that general-purpose models often overlook. The architecture utilizes a ViT-g/14 image encoder and a Vicuna LLM decoder.
2. Counterfactual Subobject Explanations (The Obfuscation Module)
The system identifies the smallest area to blur using counterfactual reasoning, asking: "What is the minimum number of image regions I need to hide to make the classifier change its mind from 'unsafe' to 'safe'?" This module employs Bayesian Adaptive Superpixel Segmentation (BASS) to break the image into meaningful shapes. Using FullGrad attribution, the algorithm ranks regions by their influence on the classifier's decision and blurs them sequentially until the image is no longer classified as unsafe.
Results and Key Insights
The framework was evaluated on three datasets: Sexually Explicit (SE), Cyberbullying (CB), and Self-Harm (SH).
- Rationale Accuracy: ConditionalVLM achieved 93.9% accuracy in categorizing unsafe images, vastly outperforming state-of-the-art models like mPLUG (47%) and InstructBLIP (50%). On specific moderator questions, it showed an average improvement of 38.2%.
- Surgical Obfuscation: The counterfactual method using BASS successfully identified and blurred unsafe regions with an 81.8% success rate. It maintained a balance where only 35-50% of the image was obscured, preserving the majority of the context.
- Segmentation Performance: Researchers found that adaptive segmentation (BASS) significantly outperformed standard methods like SLIC or SAM. BASS creates boundaries that respect actual object shapes, leading to more precise and less wasteful blurring.
Conclusion
The proposed framework offers a significant advancement in transparent and context-aware content moderation. By grounding VLMs in domain-specific classifiers and employing counterfactual reasoning for obfuscation, it provides a scalable solution for safer online environments. Future work will focus on reducing the computational complexity of the greedy search algorithm and expanding the system's ability to handle complex scenes with distributed harmful content.
Summary updated