Artificial Intelligence-Based Detection of Deepfake and Synthetic Digital Content

Abstract

Generative adversarial networks and diffusion models now synthesise faces that observers cannot separate from photographs, and the same tools manipulate identities in circulating media. Detectors that read only the visible appearance of an image inherit the weakness they are meant to expose, because modern generators are trained precisely to make appearance convincing. This work presents SynFuse-Net, a three-stream detector that reads an image simultaneously as content, as a spectrum and as a sensor noise field. A fine-tuned convolutional trunk supplies semantic evidence, a discrete cosine transform stream exposes the periodic upsampling signature left by transposed convolution and latent decoding, and a fixed high-pass residual stream isolates the local noise inconsistency produced when a face is composited into a host image. The three token sets are tagged, weighted by a learned reliability gate and fused by two blocks of cross-stream self-attention, so the contribution of each view is decided per image rather than fixed by design. It is evaluated on 190,335 manipulated and genuine faces, and on a second corpus separating genuine images, deepfakes and fully synthetic pictures. It reaches 97.44 percent accuracy and 99.69 percent area under the curve on the first corpus and 99.30 percent accuracy on the second, using 5.83 million parameters and 1.13 milliseconds per image. Four recent detectors report lower accuracy.