deepfake
A deepfake is synthetic video or audio generated to convincingly replace one person's likeness or speech with another's, and the term traces back to a specific origin: a Reddit user posting under the username "deepfakes" in late 2017, who used a shared-encoder autoencoder to swap faces in videos. That original architecture — simple by today's standards, but genuinely novel at the time — set the pattern nearly every face-swap method since has followed in some form, and it's worth understanding precisely because its structure explains both why deepfakes work and why certain early detection tricks worked against them.
The classic face-swap architecture trains one shared encoder and two separate, person-specific decoders, then swaps the decoders at inference time. The encoder learns to compress any face into a general, identity-agnostic latent representation — pose, expression, lighting — while each decoder learns to reconstruct one specific person's face from that latent code. Because the encoder is shared, feeding person A's video frames through the encoder and then through person B's decoder reconstructs the expressions and pose of A's performance, but rendered with B's face — the swap happens simply by routing the same latent code to the wrong decoder, not because the model was ever explicitly trained to do face-swapping at all.
The First Order Motion Model removed the requirement to train a new network for every new identity, which is what took face reenactment from a slow research project to something runnable on a single image. Siarohin and colleagues (2019) showed that a model could be trained once, generically, to transfer motion from a "driving" video onto any still source image without needing identity-specific training data at all — the model learns a general representation of motion as keypoint displacements and warps the source image to follow the driving video's movements. This generalization is the difference between a technique that required hours of per-person training footage, limiting it to well-documented public figures, and one that works from a single photograph, which is what made the technology accessible far beyond the original research and hobbyist communities training custom autoencoders.
Early detection relied on artifacts specific to the generation method of the moment, which is exactly why those signals had a short shelf life. Li and Lyu (2018) found that early deepfakes blinked far less often than real video, because the face datasets used to train them were dominated by open-eye photographs and rarely captured the brief closed-eye moment of a natural blink — a clean, simple detection signal, until generators simply started including blinking footage in their training data and the tell disappeared within months. This pattern repeats across the field's history: a detector finds a physical or statistical regularity real video has that synthetic video lacks, publishes it, and the next generation of generators is trained specifically to reproduce that regularity, because the published detection paper is itself a specification of what to fix.
Remote photoplethysmography detection targets something harder to fake because it isn't visible to a human eye at all. A person's pulse causes extremely subtle, rhythmic changes in skin color as blood flow varies — invisible to casual viewing but recoverable by analyzing pixel-level color variation over time. Because this signal comes from an actual physiological process rather than a visible facial feature like blinking, it's substantially harder for a generator to accidentally reproduce correctly unless it's specifically trained against this detection method too, which is part of why pulse-based detection has had a longer useful life than purely visual artifact detection — though diffusion-based generation methods, which tend to preserve more of the source video's real underlying signal than older GAN-based approaches, have begun eroding that advantage as well.
The same underlying technology generating harmful deepfakes also drives entirely legitimate film, dubbing, and accessibility applications, and no technical property distinguishes the two uses. Face reenactment and voice synthesis are used for de-aging actors, posthumous film completion, dubbing a film into another language with matching lip movement, and giving people who've lost their voice a synthetic version of how they used to sound — the exact same generation techniques, aimed at a use the subject consented to. This is why the field's policy response has increasingly focused on provenance and consent rather than banning the underlying generation technology itself: content authentication standards and cryptographic signing at the point of capture aim to establish whether a piece of media was synthetically altered and by whom, rather than trying to make the generation technology itself impossible to build.
Read deepfake generation and deepfake detection as two sides of the same adversarial research problem, not separate fields: every detection signal is a specification for what the next generator needs to fix, and every generation improvement narrows the gap detection has to find — which is why provenance and authentication, verifying what a piece of media actually is rather than trying to spot what's wrong with a fake, has become the field's more durable answer to a problem that pure after-the-fact detection structurally cannot win outright.