Catching Out-of-Context Misinformation with Self-supervised Learning

1 article/video left!

log in or sign up to unlock 3 more articles/videos this month and explore our expert resources.

Automatic Summary

Combating Misinformation: Detecting Out-of-Context Images on Social Media

Hello, everyone. I am Shivangi, a PhD student at TUM, and today I want to bring to light a rampant problem that we encounter on social media daily - misinformation. Specifically, I will be discussing one particular type of misinformation, namely out-of-context images, which I've been researching along with my advisors, Chris and Matthews.

This blog post aims to be as accessible as possible, regardless of your technical background. For a more in-depth exploration of the project, please refer to our project page linked at the end.

Why Should We Care About Detecting Out-of Context Images?

In our digital age, many of us consume most of our news and information via various social media platforms. But how do we discern which articles are true and which are false?

Fact-checking organizations serve as the answer for many companies. Yet, employing third-party fact-checkers leads to billions of dollars in company expenditure annually. With our project, we aim to automate this fact-checking process, increasing accuracy and reducing labour.

Defining the Problem

Misuses of images on social media generally fall into two categories:

  • Image Tampering: It involves playing around with or altering parts of an image to misrepresent its messaging. Multiple methods are available to detect such manipulative actions
  • Genuine Images with Irrelevant Claims: Our project focuses on this category, where an original, untampered image is shared with unrelated news or information—essentially, out of context uses of images.

A quintessential example would be an image from the 2016 anti-Trump protests circulated with a false caption about migratory riots, wherein people were burning the American Flag. Our task is to detect such misuse of images.

The Challenge of Limited Data

While available data on social media is overwhelming, the subset of images used to spread misinformation is minuscule, making it difficult to build a supervised model based on this limited training data for our problem.

Therein lies the question central our project: Can we apply the 300 million daily shared and non-misleading images on Facebook to build a model that identifies out-of-context images?

An Unsupervised Strategy for Model Training

Our project entails constructing an unsupervised or self-supervised training scheme. This involves gathering images and captions, identifying which captions are positively associated with an image, and distinguishing which captions have negative associations with that image. This process is vitally important to learning object association that ultimately helps us detect out-of-context images.

Detecting Out of Context Images: Test Time

In the testing phase, we identify relevant objects for each caption and determine a sentence similarity between the two captions.

The principle of our test strategy is simple: If the captions are semantically different but identify the same object in the image, it's likely an out-of-context use of an image. If not, it's considered in-context use. This process yields approximately 85% accuracy in detecting out-of-context images, outperforming other global context models in our comparisons.

Get Involved in the Fight Against Misinformation

If you are interested in combating digital misinformation, you might want to consider participating in the ACM Multimedia System's "Grand Challenge on Detecting Deep Fakes". They are awarding a significant prize to the challenge's winner. You can find more information and register on their website.

Feel free to reach out if you have any questions regarding this issue. Thank you for your interest and for actively participating in the fight against misinformation on social media.

Please refer to our project page for a more detailed exploration of the topic [insert project link].


Video Transcription

Read More