Abstract
Augmented Reality (AR) introduces a novel security frontier where virtual overlays can subtly alter a user's perception of the physical world. Unlike traditional "obstruction attacks" that block vision, Visual Information Manipulation (VIM) attacks create "realistic lies" by changing the semantic meaning of real-world objects. This paper introduces VIM-Sense, a multimodal semantic reasoning framework that utilizes Vision-Language Models (VLMs) and Optical Character Recognition (OCR) to detect these manipulations. By analyzing the semantic gap between raw camera feeds and AR-augmented views, VIM-Sense achieves an 88.94% detection accuracy on a newly developed AR-VIM benchmark dataset.
The Challenge of Semantic Attacks in AR
As AR technology matures, its security vulnerabilities shift from technical glitches to cognitive exploits. Traditional computer vision security often focuses on pixel-level anomalies or physical realism (e.g., lighting and shadows). However, VIM attacks are designed to be stylistically consistent with the environment, making them nearly impossible for users to distinguish from legitimate information.
In a typical VIM scenario, a malicious application might change a speed limit sign or alter a directional arrow. Because the human brain is wired to trust its visual perception, these subtle changes can lead to dangerous real-world consequences. Detecting these attacks requires a system that understands context and intent, rather than just visual consistency.
The VIM-Sense Framework
To address this "semantic gap," the researchers developed VIM-Sense, which operates on an edge-cloud architecture. The system follows a four-step multimodal reasoning process:
- OCR Grounding: The system performs OCR on both the raw camera feed and the AR view to extract literal text, providing a factual anchor for the model.
- Chain-of-Thought Prompting: A Vision-Language Model (such as GPT-4o) is guided through a reasoning process to identify the real-world scenario and the nature of the added virtual content.
- Multimodal Reasoning: By comparing the images and the extracted text, the VLM identifies whether the AR content alters the original meaning of the scene.
- Cognitive Resilience Modeling: The system is instructed to ignore obvious or stylistically inconsistent virtual objects, focusing specifically on subtle, deceptive manipulations.
The AR-VIM Dataset
To evaluate VIM-Sense, the authors created the AR-VIM Dataset, comprising 452 raw-AR video pairs across 202 unique scenes. The dataset covers seven types of attacks, ranging from simple character replacements to complex phrase alterations and misleading patterns.
The dataset was validated through a user study with 26 participants, confirming that the attacks were perceived as misleading by humans. This benchmark allows for a rigorous testing of how well AI can protect users from cognitive deception in AR environments.
Key Results and Insights
VIM-Sense demonstrated significant efficacy, particularly when using high-level models like GPT-4o. Key findings include:
- Superior Accuracy: The framework achieved 88.94% accuracy, with a peak of 95.16% for phrase replacement attacks.
- OCR as a Critical Anchor: The study found that VLMs alone are prone to hallucinations. Explicitly providing OCR text prevents the model from misreading subtle character changes (e.g., misidentifying a '3' as an '8').
- Prompt Engineering: Instructing the model to assume the user is "not easily fooled" reduced false positives, preventing the system from flagging harmless virtual objects as attacks.
- Latency: The system currently operates with an end-to-end latency of approximately 7.17 seconds. While sufficient for stationary tasks like reading signs, further optimization is required for dynamic, high-speed scenarios.
Conclusion and Future Work
VIM-Sense represents a significant step toward securing the cognitive layer of Augmented Reality. By treating AR security as a semantic reasoning problem rather than a purely visual one, the framework can identify sophisticated "lies" that would otherwise bypass traditional security measures. Future research will focus on reducing latency through lightweight local models and expanding the scope of detection to include manipulated spatial audio and haptic feedback.
Summary updated