Page 16 - Detecting deepfakes and generative AI: Report on standards for AI watermarking and multimedia authenticity workshop

Detecting deepfakes and generative AI: Report on standards for AI
watermarking and multimedia authenticity workshop

Existing deepfake detection tools can be broadly categorized into two groups:

i) Based on either the exploitation of semantic inconsistencies like irregular eye reflections or
known generation artifacts in the spatial or frequency domain.
ii) Using neural networks to learn a feature representation in which real images can be
distinguished from AI-generated ones. For instance, training a standard convolutional
neural network (CNN) on real and fake images from a single GAN yields a classifier
capable of detecting images generated by a variety of unknown GANs. Given the rapid
evolution of generative AI models, developing detectors which generalize to new
generative AI models is crucial and therefore a major field of research.

Deepfake detection techniques use deep learning and machine learning to analyse patterns
and anomalies in multimedia content and identify signs of manipulation. Detection techniques
can be split into two detection methods: CNN-based methods and region-based convolutional
neural networks (R-CNN)-based methods. CNN-based techniques take pictures of people's
faces from video frames and feed them into the CNN for training and prediction to get an
image-level result. Such algorithms only employ spatial information from a single frame.

R-CNN-based techniques, on the other hand, require a series of video frames for training to
produce a video-level result. This approach, known as R-CNN, combines CNN and recurrent
neural networks (RNN). As a result, R-CNN-based techniques could fully use spatial and
temporal information in deepfake video.

In addition, several deepfake detection methods are based on standard machine learning
methods, including utilizing a support vector machine (SVM) as a classifier and extracting
handmade characteristics, including biological signals, as a classifier. For example, the video
of a person’s face contains subtle shifts in colour that result from pulses in blood circulation.
These changes in colour form the basis of a technique called photoplethysmography (PPG)
that can be used to detect synthetic media. Deepfakes cannot recreate these colour shifts with
high fidelity. Biological signals are not coherently preserved in different synthetic facial parts
and deepfakes do not contain frames with stable PPG.

Current deepfake video detection methods have several limitations. Firstly, these methods
cannot always be relied upon to detect deepfakes in real-world situations, especially when the
images or videos are modified using new techniques that were not present in the training data.
Most methods fail to model the natural structures and movements of human faces adequately,
which are crucial for accurate deepfake detection. Some methods rely heavily on mouth-related
mismatches between auditory and visual modalities, leading to performance degradation when
there are limited or unaltered mouth motions. These limitations highlight the need for improved
deepfake detection methods that can effectively handle real-world scenarios, generalize well
to unseen samples, and capture the natural cues of human faces.

Touradj Ebrahimi, Professor at EPFL and Chair of JPEG, presented a new framework for
deepfake detection in still images that could enhance the performance of a deepfake detector
under the attack of various real-world perturbations (e.g., JPEG compression artifacts, changing
brightness and contrast, blurry effects, and Gaussian noise). He presented two methods:

a) Stochastic degradation-based augmentation.
b) Degradation-based amplitude-phase switch augmentation.

He concluded by presenting a detection technique that allows for the detection of content
synthesized completely by generative AI techniques.

8