None of the 16 leading deepfake detectors can reliably identify fake photos in real-world conditions, Australian and South Korean researchers have found.
Research reveals serious weaknesses in technology
According to a newly published study on the arXiv portal, jointly conducted by the Australian national science agency CSIRO and South Korea's Sungkyunkwan University, existing deepfake detectors show serious vulnerabilities.
The research analyzed 16 leading deepfake detection algorithms and showed that none of them could detect fake images with sufficient accuracy in real-world scenarios.
The researchers developed a five-step methodological approach to evaluating detectors, which includes:
• The type of deepfake
• Detection method
• Data preparation
• Model training
• Validation of results
Additionally, they identified 18 factors that affect the accuracy of the detectors and tested them in different scenarios, including "black box", "white box", and "gray box" (different levels of access to training data).
The detectors failed on various types of deepfake content
Existing deepfake detectors show significant weaknesses, especially when faced with content that was not included in their training set.
For example, “ICT” (Identity Consistent Transformer) – a model trained on faces of famous people, failed to detect deepfakes in unknown people.
🔹 "Synthesis deepfake" – Generates completely new synthetic faces
🔹 "Faceswap deepfake" – Swaps one person's face with another
🔹 "Reenactment deepfake" – Keeps the face, but changes expressions and movements
Regardless of type, detectors have failed to achieve high accuracy.
Urgent improvements are needed.
The researchers call for urgent improvements in detection technologies. They recommend:
• ️ Development of more advanced detectors
• ️ Using a wider range of training data
• ️ Integration of audio, text and metadata
Additionally, new strategies have been proposed, such as:
• Fingerprinting – Embedding artificial traces in images and videos, thereby tracing their origin
• GAN fingerprints – Recognition of natural features left by generative models
Existing tools are not effective enough
This research shows that current deepfake detectors are not accurate enough for real-world applications.
Given the rapid advancement of artificial intelligence, innovative methods and more efficient models are needed to successfully recognize and combat digital counterfeits.
Research reveals serious weaknesses in technology
According to a newly published study on the arXiv portal, jointly conducted by the Australian national science agency CSIRO and South Korea's Sungkyunkwan University, existing deepfake detectors show serious vulnerabilities.
The research analyzed 16 leading deepfake detection algorithms and showed that none of them could detect fake images with sufficient accuracy in real-world scenarios.
The researchers developed a five-step methodological approach to evaluating detectors, which includes:
• The type of deepfake
• Detection method
• Data preparation
• Model training
• Validation of results
Additionally, they identified 18 factors that affect the accuracy of the detectors and tested them in different scenarios, including "black box", "white box", and "gray box" (different levels of access to training data).
The detectors failed on various types of deepfake content
Existing deepfake detectors show significant weaknesses, especially when faced with content that was not included in their training set.
For example, “ICT” (Identity Consistent Transformer) – a model trained on faces of famous people, failed to detect deepfakes in unknown people.
🔹 "Synthesis deepfake" – Generates completely new synthetic faces
🔹 "Faceswap deepfake" – Swaps one person's face with another
🔹 "Reenactment deepfake" – Keeps the face, but changes expressions and movements
Regardless of type, detectors have failed to achieve high accuracy.
Urgent improvements are needed.
The researchers call for urgent improvements in detection technologies. They recommend:
• ️ Development of more advanced detectors
• ️ Using a wider range of training data
• ️ Integration of audio, text and metadata
Additionally, new strategies have been proposed, such as:
• Fingerprinting – Embedding artificial traces in images and videos, thereby tracing their origin
• GAN fingerprints – Recognition of natural features left by generative models
Existing tools are not effective enough
This research shows that current deepfake detectors are not accurate enough for real-world applications.
Given the rapid advancement of artificial intelligence, innovative methods and more efficient models are needed to successfully recognize and combat digital counterfeits.
