Quantized Contour-Based Intelligent Stegomalware Sterilizer for Smart Consumer Electronics Network

Malware attacks on consumer electronics networks often originate from harmful codes concealed within seemingly benign images acquired from outside the consumer network. Due to limited content inspection in consumer firewalls, malware concealed within image infiltrates into consumer network. To address this concern, this work presents an image sterilization framework that eliminates hidden payloads via sterilization network and recovers image quality through repair network. The sterilization network constructs a quantized stego-contour to quantify the likelihood of pixel-wise payload-carrying capacity, and utilizes it for performing removal attack on the Catalan coefficients of the suspicious image. The repair network, on the other hand, restores the image from visible distortions resulting from sterilization attack. Unlike traditional holistic image modification paradigms, sterilization via quantized stego-contour limits unnecessary perturbations and eliminates the need for prior knowledge regarding embedment. Pixel-level spatial sterilization tends to produce statistical discontinuities, attacking Catalan coefficients facilitates localized noise dispersion ensuring minimized distortion. The effectiveness of our framework is evaluated against script-based and image-based embedding algorithms from robust steganography, watermarking, and deep data hiding categories, using benchmark datasets, i.e., CelebA and MS COCO. Experimental results demonstrate that the proposed sterilizer achieves average PSNR of 48.40 dB with removal rate of 0.95 against the GMAS embedding on CelebA and obtains removal rate of 0.94 on real malware payload with throughput of 21 images per second. The proposed framework outperforms state-of-the-art sterilizers, i.e., PixelSteganalysis, ScNet, EBRA, and GAF and provides stable outcome against distributional shifts, i.e., compression, cropping establishing its effectiveness for real-time image sanitization.