Quantifying JPG Ugliness to Fight Disinformation
Greetings!
If you know my work, you know that I’m deeply invested in research methods drawn from digital forensics. I like to pull things apart, reverse-engineer them, and use that process to learn something new about how they came to be.
I just co-authored a paper (and software package) applying that approach to the study of hoaxes, fake news, and disinfo online. Basically, we saw a bunch of sketchy images being used to circulate misleading narratives around the January 6 Capitol Insurrection, and we wanted to add more depth to the story using a forensic analysis of those images.
As it turns out, you can use the ugly JPG compression artifacts in an image as a thumbprint that identifies (roughly) how much it has been shared and re-shared. When you combine that quantitative analysis with a more interpretive analysis of all the weird paratextual elements that come from people screenshotting and cropping the image, you can actually get a pretty good idea of how it spread through different networks over time. Using this approach, we were able to identify a network of coordinated accounts pushing false narratives. Pretty cool!
This project should be of interest to anyone looking into disinformation, image compression, or digital forensics. It’s my first foray into co-authorship, as well as my first time working on original software. It’s also a neat combination of computational and interpretive methods, for anyone else who likes to bridge the STEM-humanities gap (I know I do).
The paper is out now in the proceedings for the upcoming American Society for Information Science & Technology (ASIS&T) conference (co-authored with Mitch Chaiet of Google and Praful Gupta from Amazon). We called the new method “SMOC BRISQUEt” because it builds on the existing BRISQUE image analysis algorithm, and because smoked brisket is a Texan BBQ specialty (Mitch and Praful were both working at UT when we wrote this paper, although they’ve since moved on to new gigs). This is my first time using a pun in an academic paper.
I also recorded a video lecture for the conference, which will debut next week. You’ll have to be at the event in real time to catch the video for now, but I’m happy to share it on a one-to-one basis if you have any interest. Just get in touch!
Links:
“Forensic Analysis of Memetic Image Propagation: Introducing the SMOC BRISQUEt Method.” [link]
SMOC-BRISQUEt on GitHub [link]