AI-Enhanced Mobile Diminished Reality for Preserving 3D Visual Privacy
Salam Tabet, Ayman Kayssi, Imad H. Elhajj · 2024
In today's digital society, the concept of privacy is rapidly diminishing. The proliferation of devices that record and monitor personal information, especially with the rise of immersive technologies, exacerbates this issue. Existing privacy-preserving techniques are inadequate for three-dimensional environments, often failing to maintain performance during significant camera motion or when balancing privacy with utility. Our previous work introduced a three-dimensional visual privacy-preserving Diminished Reality (DR) framework for mobile devices. We also proposed a Utility-Privacy Tradeoff Algorithm (UPTA) that achieves a proper balance between privacy and utility, where utility is defined as a geometric measure that is inversely proportional to the size of our obfuscation. But since any exposure of the private object, even within a single frame, can greatly violate the user's privacy, in this paper, we propose an enhancement to this framework using a Machine Learning (ML) model, referred to as AI-UPTA. This model predicts the potential failure of UPTA in real time, ensuring complete screen obfuscation when necessary to prevent privacy leaks. AI-UPTA evaluates detected objects, camera movements, and environmental factors to improve privacy without significantly compromising utility. Our evaluation demonstrates that AI-UPTA effectively increases privacy levels, compared to state-of-the-art, while maintaining an acceptable utility tradeoff, improving our baseline system by 54%.