Threats Caused by AI and Overcoming Them: A Look at the Cases of DALL-E2 Harmful Content Generation and DeepFake Videos Detection

Description

Speaker: Elif Bilge Kalvun, University of Passau, Germany. 

Hosted by Pufendorf Advanced Study Group: Deep Generative Modeling as well as by the eScience Hub at Lund University.
 

Abstract: With the successful implementation of Artificial Intelligence and specifically Generative Adversary Networks applications in the last few years, digital manipulations to images and faces with Deep Learning has become popular among researchers and social media users. In this talk, we will discuss the identity security and privacy threats caused by AI in two specific cases.

In the first example, we will look into DALL-E2, which is a state-of-the-art AI model developed by OpenAI that generates unique images from text inputs and allows for image editing. The increasing security concerns surrounding DALL-E2 highlights the potential for circumvention of the model's filters for banned words to create harmful content together with its editing features and the ease of using it to attack individuals and/or their privacy. We will discuss these problems together with potential mitigation methods.

Secondly, we will look into DeepFake, which is a well-known and promising face manipulation technology to modify and swap individuals' faces in videos. DeepFake technology has originated some concerns regarding personal privacy and identity, which can be used in forgery, defamation, spreading disinformation, etc. As a countermeasure against DeepFakes, we focus on the rapid and accurate detection of manipulated videos by utilizing transfer and ensemble learning with the combination of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). Our investigations so far show that detecting manipulated videos is possible, and we can improve the previously reported AUC scores by combining different trained models (EfficientNetB4 + Xception) and reach a 99% AUC score.