Digital image forensics includes methods for verifying the authenticity of digital images, attributing images to potential sources, and detecting traces of manipulation. Many images subject to forensic analyses are compressed with JPEG. During JPEG compression, the Discrete Cosine Transform (DCT) converts blocks of 8x8 pixels into a frequency domain representation. If blocks contain combinations of pixels that cannot be exactly represented by the discretized cosine functions, the decompressed image will exhibit visual artifacts. Known as ringing artifacts, they typically appear around sharp edges or text and are visible as halos or a faint glow. Mozjpeg, a popular JPEG compression library, implements an algorithm that modifies the frequency domain representation of the blocks in order to avoid ringing artifacts.
The objective of this thesis is to analyze the deringing algorithm and evaluate its impact on the field of digital image forensics. For example, to what extent are source attribution (e.g., to a pre-processing history) and manipulation detection (e.g., traces of redacted content) affected by the use of deringing? To what extent does ringing (and thus deringing) interfere with JPEG steganalysis? Can we possibly design a steganographic scheme that imperceptibly hides information in the ringing artifacts? Or one that uses deringing to identify more secure hiding positions within the image as in “natural” steganography?
Under the guidance of the supervisor, the student will start with a literature review of forensic methods that are susceptible to JPEG ringing, create or adapt benchmark image data to experimentally measure the effect, propose and finally evaluate possible mitigation strategies. We expect the student to focus on one of the questions outlined above.