Forensic Recognition of Codec-Specific Image Compression Artefacts

Forensic Recognition of Codec-Specific Image Compression Artefacts

연구 분야: Strategies

학회: IH&MMSec '24: Proceedings of the 2024 ACM Workshop on Information Hiding and Multimedia Security

This work investigates the possibility to conduct a forensic discrimination of decoded versions of 10 different lossy image compression file formats, including 4 ISO/IEC still image compression standards (JPEG, JPEG 2000, JPEG XR, JPEG XL) and 4 video-coding related image compression schemes (AVIF, HEIC, BPG, WEBP). We have found that a proper compression artefact discrimination can be achieved across different compression ratios by fine-tuning a standard ResNet-18 model using a variety of different file sizes in training. Classification accuracy is almost perfect for low quality image data (as compression artefacts are strong), while the 10-class discrimination accuracy is slightly beyond 85% for high quality imagery which can be considered almost visually lossless. Observed mis-classifications are mostly along the lines of expectations due to algorithmic differences and similarities (block-size, transform type, etc.), only JPEG 2000 exhibits some unexpected artefact similarities to JPEG XR when the photo overlap transform is being employed.