What can we detect?
Synthetic identities are commonly generated in form of images by GANs (Generative Adversarial Networks). GANs used as fake personas and bot accounts on social media and dating sites. Sensity's detector are trained on millions of GAN-generated images found online under various conditions of compression, cropping, and photo filters. They are trained to spot artefacts and high- frequency signals characteristics of AI- generated images, which natural photos are unlikely to possess.
Detecting DALL-E, Stable Diffusion and Mid Journey
Diffusion models like Dall-E, Stable Diffusion, Mid Journey are extraordinary tools for content creation without being a graphic expert. The high accuracy in creating realistic face and full body figures will definitely pose challenges to detection solutions. Sensity can detect diffusion models creation with the 95.8% accuracy.
Deepfakes are used in realtime by fraudsters to impersonate their victims in the case of identity theft. Face swap and face reenactment are similar techniques where a target victim's face is either swapped onto the attacker's head, or controlled by the attacker's head movements. Similarly, deepfakes are also used to generate thousands of fake personas that can move and talk, with the goal to spoof automated KYC processes by brute force. Deepfake detection uses the latest advances in deep learning and image forensics to determine whether a person in the video is genuine.
AI Generated Text
AI generated text by Large Language Models such as GPT-3 has become indistinguishable from genuine human text. These tools are exploited for cheating during online, plagiarism, and for automating misinformation, social engineering and cyber attacks.
Zero-shot machine-generated detection is achieved by evaluating the probability that chunks of text are produced by popular AI text services. The detector is robust to content editing by human writers, such as changes of words, punctuation and also misspellings.