The human eye can distinguish between authentic and fake images by observing factors such as the number of fingers in the image, the strangeness of the eyes, and whether the hair looks real. Generated photos often exhibit flaws in lighting and shadow, lacking basic parameters such as exposure. Additionally, the clarity of teeth can be a clue in identifying deepfake images, with the absence of individual tooth contours potentially indicating manipulation.A more efficient method involves AI combating AI. OpenAI has announced the launch of a tool designed to detect images created by the AI image generator DALL-E 3. Microsoft has developed an authentication tool capable of analyzing photos or videos to determine if they have been manipulated.
As deepfake technology becomes increasingly accessible, distinguishing between real and fake becomes more challenging.
Tools like DALL-E, Midjourney, and Sora can generate images and videos with a simple request, making deepfake creation more accessible. These deceptive images, though seemingly harmless, can be used for fraud and manipulation, such as during the Indian general elections in April, when fake videos critical of Prime Minister Modi circulated online.
How to Identify Deepfakes? The Associated Press Provides Several Methods
In the early stages of deepfake technology, imperfections often left manipulation traces, such as hands with six fingers or discrepancies in the shape of eyeglass lenses. However, as AI progresses, these observations become more challenging. Henry Ajder, founder of Latent Space Advisory, notes that some widely circulated suggestions, like identifying unnatural blinking patterns in deepfake videos, are no longer reliable, though other signs remain detectable. Certain AI-generated deepfake photos, especially of people, exhibit an electronic sheen and a “smooth aesthetic effect” that makes skin appear “extremely smooth.” However, Ajder also suggests that creative hints can sometimes eliminate this smoothness and other signs of AI manipulation.
According to Latest report,. Lyu Qiang, an algorithm scientist at XinYe Technology, suggests that discerning synthetic images with the naked eye requires a focus on details such as the normalcy of finger count, the strangeness of eye expressions, and the authenticity of hair strands. Additionally, generated photos often lack flaws in lighting and shadow, as well as basic parameters like exposure.
Consistency in shadows and lighting is one identification method; while the focus on the subject of a photo may be clear and lifelike, elements in the background may appear less realistic or altered. Exaggerated or out-of-character actions by public figures in an image may also indicate a deepfake. Furthermore, facial swapping, a common method of deepfake creation, can be detected by closely examining the edges of the face to see if the skin tone matches other parts of the head or body. In videos, suspicions of deepfakes can arise if lip movements do not perfectly match the audio. Ajder advises observing the clarity of teeth and their consistency with real-life appearances. Norton, a cybersecurity company, suggests that algorithms may not be sophisticated enough to generate individual teeth, making the absence of tooth contours a clue in identifying deepfakes.
Currently, Microsoft has developed an authentication tool that can analyze photos or videos and provide ratings on whether they have been manipulated. OpenAI recently announced the forthcoming release of a tool designed to detect images created by the AI image generator DALL-E 3. OpenAI reports that in internal testing, this tool correctly identified DALL-E 3-generated images 98% of the time and was minimally affected by common modifications such as compression, cropping, and changes in saturation. In addition, OpenAI has joined an industry organization including Google, Microsoft, and Adobe and plans to provide standards to help track various media sources. Intel’s FakeCatcher uses algorithms to analyze image pixels to determine authenticity. According to CNN’s report in February of this year, Meta is also working on identifying and labeling AI-generated images shared on its platform by third-party tools.
In China, SenseTime’s digital watermarking technology can embed specific information into multimodal digital carriers, including images, videos, audio, and text, without affecting the quality or perception of the original content. This technology can withstand attacks to prevent deletion or modification. SenseTime indicates that digital watermarking technology can be used for scenarios such as copyright protection and anti-counterfeiting. It can maintain over 99% watermark extraction accuracy under different degrees of interference such as cropping and compression without compromising data quality. As the technology landscape evolves, a multi-pronged approach integrating technological advancements, collaborative efforts, and regulatory frameworks will be crucial in effectively combating deepfakes and preserving trust in digital media.
