[Paper Review] Block-wise Scrambled Image Recognition Using Adaptation Network

Koki Madono, Masayuki Tanaka+2 more|arXiv (Cornell University)|Jan 21, 2020

Digital Media Forensic Detection13 references31 citations

TL;DR

The paper proposes an adaptation network to recognize block-wise scrambled images produced by extended learnable encryption (ELE), enabling perceptual information hiding with DNN-based classification. ELE-based scrambling paired with ELE-AdaptNet yields reliable recognition under scramble, while LE-based scrambling remains less secure.

ABSTRACT

In this study, a perceptually hidden object-recognition method is investigated to generate secure images recognizable by humans but not machines. Hence, both the perceptual information hiding and the corresponding object recognition methods should be developed. Block-wise image scrambling is introduced to hide perceptual information from a third party. In addition, an adaptation network is proposed to recognize those scrambled images. Experimental comparisons conducted using CIFAR datasets demonstrated that the proposed adaptation network performed well in incorporating simple perceptual information hiding into DNN-based image classification.

Motivation & Objective

• Motivate privacy-preserving cloud-based image analysis by hiding perceptual information from humans while allowing machine recognition.
• Extend learnable encryption (LE) to strengthen security of scrambled images.
• Develop an adaptation network capable of recognizing block-wise scrambled images produced by ELE/LE/EtC schemes.
• Compare recognition accuracy across plain, LE, EtC, and ELE scrambling on CIFAR-10/100 datasets.
• Provide a mechanism to balance security level and recognition performance for scrambled images.

Proposed method

• Introduce block-wise scrambling where images are divided into B×B blocks, blocks are shuffled and pixels within blocks are shuffled with per-block keys (ELE).
• Propose an adaptation network consisting of block-wise sub-networks, a learnable pseudo permutation matrix, and a pixel shuffling layer.
• Each scrambled block is processed by a dedicated sub-network; features are integrated, passed through a pseudo permutation alignment, upscaled via pixel shuffle, and fed to a classification network.
• Train with a composite loss: cross-entropy for classification, L1-2 penalty to encourage sparse pseudo-permutation, and a spatial-smoothness penalty on the feature maps.
• Loss: L = L_CE + λ_U L_U + λ_s L_s with empirically chosen λ_U = 0.001 and λ_s = 0.1.

Experimental results

Research questions

• RQ1 Can a single adaptation network recognize block-wise scrambled images across multiple scrambling schemes (LE, EtC, ELE)?
• RQ2 How does recognition performance trade off with security level across different scrambling schemes?
• RQ3 Does introducing a pseudo permutation matrix and spatial regularization improve robustness to block-wise scrambling during classification?
• RQ4 What is the comparative performance of plain versus scrambled images in CIFAR-10/100 under the proposed adaptation framework?

Key findings

• ELE-AdaptNet achieves the best recognition performance for ELE- and EtC-based scrambling.
• LE-AdaptNet performs best for LE-based scrambling due to specialization to LE.
• Plain images yield high accuracy and scrambled images degrade accuracy, with ELE providing a favorable balance between security and recognition under the proposed adaptation network.
• ELE-based scrambling paired with the proposed adaptation network yields reliable posterior probabilities even when scrambling is strong, indicating robust recognition under perceptual hiding.
• The ELE method provides a larger key space than LE or EtC, indicating higher security for perceptual information hiding while remaining compatible with the adaptation network.