Synthetic Media Detection as a Security Problem: A Cross-Modal Survey of Methods, Benchmarks, and Robustness Limits

Abstract

Generative models now produce faces, voices, video and text that people cannot reliably tell from real recordings, and the losses are no longer hypothetical: one engineering firm wired about US$25 million to criminals after a video call in which every other participant was synthetic. This paper surveys the detection side of the problem across four media types. We organize visual detectors by the signal they exploit (spatial artifacts, frequency artifacts, temporal and biological cues, and features from large pretrained vision models), trace audio detection from the ASVspoof challenges to self-supervised front-ends, and cover text and audio-visual detectors more briefly. We then collect the published numbers that matter for deployment. Detectors that reach 96 to 99.9% AUC on the dataset they were trained on fall to 65 to 74% AUC on a second deepfake dataset and to chance on images from an unseen generator family. Audio detectors at about 1% equal error rate in the lab reach 34% on in-the-wild recordings. White-box adversarial perturbations drive a 0.95AUC image detector to 0.0005, and text detectors flag 61% of essays by non-native English writers as machine-written. We close with the shift toward provenance and watermarking mandated by the EU AI Act, and list the open problems: generalization to new generators, robustness under compression and attack, localized and partial forgeries, and evaluation protocols that stop detectors from learning shortcuts.