Deepfake
A deepfake is synthetic media in which a person's likeness or voice is fabricated or altered using deep learning, typically generative adversarial networks or diffusion models, to depict events that never occurred.
A deepfake is a piece of synthetic media in which an individual's face, body, or voice is generated or manipulated by deep learning techniques to make them appear to say or do things they never did. The term combines "deep learning" and "fake" and entered common usage around 2017. Deepfakes are most commonly produced using generative adversarial networks, autoencoders, and, more recently, diffusion models, which can synthesise highly convincing video and audio from relatively modest amounts of source material.
How deepfakes are made
Early face-swap deepfakes relied on a pair of autoencoders that learned to compress and reconstruct the faces of two people, then swapped the decoders so that one person's expressions were rendered with another's appearance. Generative adversarial networks improved realism by pitting a generator that creates fake frames against a discriminator that tries to detect them, driving both toward higher fidelity. Modern pipelines increasingly use diffusion models and large multimodal systems that can generate or edit video directly from text or reference images.
Audio deepfakes, often called voice cloning, use neural text-to-speech and voice-conversion models that can reproduce a target speaker's timbre and intonation from a few seconds of recorded speech. The convergence of video synthesis, voice cloning, and freely available data-scraping tools means that convincing fakes can now be produced in minutes rather than requiring specialist expertise.
Detection and provenance
Deepfake detection is an active research area, but it remains an adversarial problem: as generators improve, the visual and acoustic artefacts that detectors rely on become harder to find. Detection approaches include classifiers trained to spot statistical traces of synthesis, analysis of physiological signals such as blinking or pulse, and inconsistencies in lighting and reflections. Because detection alone is fragile, attention has shifted toward provenance and authenticity standards, such as content credentials and cryptographic watermarking that label media at the point of creation. The table below contrasts the two strategies.
| Strategy | Approach | Limitation | | --- | --- | --- | | Detection | Classify media as real or fake after the fact | Degrades as generators improve | | Provenance | Attach verifiable origin data at creation | Requires broad adoption by platforms |
Uses and harms
Deepfakes have legitimate applications in film dubbing, accessibility, satire, education, and synthetic data generation. However, they are widely associated with harm, including non-consensual imagery, political disinformation, and financial fraud in which fabricated executives or public figures authorise transfers or promote fake investment schemes. The erosion of the assumption that recorded media is authentic, sometimes called the "liar's dividend," also lets genuine evidence be dismissed as fabricated.
References
- Condition Zebra. (2025). AI Deepfake Scams in Malaysia 2025.
- Penang Institute. (2025). Urgent Need to Strengthen Malaysia's Legal Framework Against AI-Driven Scams.
- BERNAMA. (2025). Malaysia Must Strengthen AI Policies, Regulatory Framework To Curb Deepfake Scams.
- Goodfellow, I., et al. (2014). Generative Adversarial Networks. NeurIPS.