Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually (2401.11035v1)

Abstract: Social media platforms are being increasingly used by malicious actors to share unsafe content, such as images depicting sexual activity, cyberbullying, and self-harm. Consequently, major platforms use AI and human moderation to obfuscate such images to make them safer. Two critical needs for obfuscating unsafe images is that an accurate rationale for obfuscating image regions must be provided, and the sensitive regions should be obfuscated (\textit{e.g.} blurring) for users' safety. This process involves addressing two key problems: (1) the reason for obfuscating unsafe images demands the platform to provide an accurate rationale that must be grounded in unsafe image-specific attributes, and (2) the unsafe regions in the image must be minimally obfuscated while still depicting the safe regions. In this work, we address these key issues by first performing visual reasoning by designing a visual reasoning model (VLM) conditioned on pre-trained unsafe image classifiers to provide an accurate rationale grounded in unsafe image attributes, and then proposing a counterfactual explanation algorithm that minimally identifies and obfuscates unsafe regions for safe viewing, by first utilizing an unsafe image classifier attribution matrix to guide segmentation for a more optimal subregion segmentation followed by an informed greedy search to determine the minimum number of subregions required to modify the classifier's output based on attribution score. Extensive experiments on uncurated data from social networks emphasize the efficacy of our proposed method. We make our code available at: https://github.com/SecureAIAutonomyLab/ConditionalVLM

• The paper proposes a dual-module framework that integrates ConditionalVLM for classification and counterfactual techniques for precise obfuscation of unsafe regions.
• The method uses Bayesian Adaptive Superpixel Segmentation and FullGrad attribution to segment and target sensitive image areas while preserving safe content.
• Experimental results show enhanced detection accuracy and comprehensive rationale generation compared to models like mPLUG and OFA-Large.

Image Safeguarding Using Conditional Vision LLMs

Introduction

The paper "Image Safeguarding: Reasoning with Conditional Vision LLM and Obfuscating Unsafe Content Counterfactually" addresses the critical issue of obfuscating unsafe images shared on social media platforms. This work introduces a dual-module approach to ensure both effective obfuscation of unsafe regions and the generation of accurate rationales for these actions. The proposed method leverages the Conditional Vision LLM (ConditionalVLM) for reasoning and uses counterfactual visual explanations to determine and obfuscate sensitive regions within images.

Conditional Vision-LLM

The ConditionalVLM forms the first module of the proposed architecture, designed to classify images as safe or unsafe by integrating visual reasoning with conditional inspection.

Architecture

The ConditionalVLM employs a pre-trained vision transformer (ViT) for extracting visual embeddings, which are further processed through a Conditional Image Instruction-guided Transformer (CIIT). This step involves conditioning the model outputs based on unsafe image classifiers, thereby enhancing the model's ability to accurately generate rationales grounded in specific unsafe image features.

Figure 1: Overview of the proposed architecture. The initial module utilizes ConditionalVLM for classifying images as safe or unsafe, while the subsequent module proposes counterfactual visual explanations to identify and obfuscate the unsafe regions within the image.

Counterfactual Visual Explanations

The second module uses a counterfactual approach to identify and obfuscate unsafe regions in images. This technique minimizes obfuscation of safe areas, ensuring only sensitive portions are obscured, which is crucial for maintaining the utility of images during investigations.

Methodology

The process involves segmenting the image into subregions using Bayesian Adaptive Superpixel Segmentation (BASS), informed by the classifier's output. The segmentation process is guided by attribution maps from FullGrad to prioritize regions that most significantly contribute to the classification outcome. Subsequently, a discrete optimization approach using informed greedy search determines the minimal subregions that require obfuscation to change the classifier's unsafe verdict to safe.

Figure 2: Examples of segmentation methods on a cyberbullying image. From top to bottom: (1) BASS, (2) SLIC, (3) SAM.

Experimental Evaluation

The efficacy of the ConditionalVLM and the obfuscation technique was validated on datasets encompassing various categories of unsafe content, such as nudity and cyberbullying. ConditionalBLIP, an adaptation of InstructBLIP integrating CIIT, outperformed existing models in accurately generating descriptions for unsafe images. It achieved notable improvements in detection accuracy and provided comprehensive explanations of unsafe features when compared to other models like mPLUG and OFA-Large.

Counterfactual Evaluation

Counterfactual subobject explanations demonstrated significant success rates in obfuscating unsafe image regions while preserving the integrity of safe areas. The method showcased a notable enhancement over previous approaches such as CSRA, with improved success rates under realistic constraints in testing scenarios.

Conclusion

This paper contributes to advancing image safeguarding techniques on social media by introducing a robust architecture that combines conditional reasoning with precise obfuscation strategies. ConditionalVLM and counterfactual analysis not only enhance the ability to obfuscate sensitive content efficiently but also ensure the rationale provided is grounded in domain-specific unsafe attributes. These innovations ultimately support safer content sharing and aid in exposure reduction for content moderators and investigative use cases. The proposed methodology has the potential to be integrated into broader AI-driven moderation systems across various platforms.