TITLE: Exploration of Thumbnail Preserving Encryption(TPE)'s vulnerability to Super Resolution Networks

ABSTRACT: Thumbnail Preserving Encryption (TPE) was proposed to balance usability and privacy when storing images with third-party cloud providers. However, TPE images are vulnerable to super-resolution models that exploit the preserved visual structure to recover the original plaintext image. In this work, we investigate the effectiveness of a state-of-the-art super-resolution attack on TPE called AttackTPE (uses a pix2pix model). We propose an improved super-resolution attack on TPE leveraging SwinIR, a convolution transformer hybrid. We show that our SwinIR attack outperforms the state-of-the art by achieving +7.38 dB (+39.3%) Peak Signal to Noise Ratio(PSNR) and +0.3068 (+70.3%) Structural Similarity Index Measure(SSIM) over the baseline. We propose a cosine similarity metric based on facial recognition embeddings to assess the semantic similarity between the original and reconstructed images. Furthermore, we employ this metric to evaluate attack success rates and empirically determine the TPE block parameter lower bound necessary for resistance against super-resolution attacks. Specifically, we find that the TPE block area to face area ratio must exceed 1/25 for TPE to resist super-resolution attacks on portrait images.

MAJOR ADVISOR: Rakesh Bobba
COMMITTEE: Dave Nevin
MINOR ADVISOR: Sanghyun Hong
GCR: J Nicole Von Germeten