The Emergence of Deepfake Technology: A Review¶

Venue: Technology Innovation Management Review, Vol. 9, No. 11 (November 2019), pp. 39–52

TL;DR¶

This review analyzes 84 news articles about deepfakes to synthesize understanding of synthetic video technology, its applications, threats, and countermeasures. Deepfakes—hyper-realistic videos created using AI (primarily GANs)—lower the barrier to video forgery and represent a significant threat to trust, elections, and security. Solutions span legislation, corporate policy, education, and anti-deepfake detection technology, but no single approach is sufficient; a layered response is required.

Contributions¶

Comprehensive synthesis of deepfake technology from news media discourse (84 articles from 11 major news outlets)
Clear definition and taxonomy of deepfakes, distinguishing them from other forms of media manipulation
Documentation of legitimate use cases (entertainment, education, accessibility)
Enumeration of threat vectors: misinformation, national security, citizen trust, cybersecurity, and fraud
Survey of four complementary defense strategies: legislation, corporate policy, education, and anti-deepfake technology

What are Deepfakes?¶

Deepfakes are hyper-realistic synthetic videos created using machine learning—specifically, Generative Adversarial Networks (GANs)—that depict people saying or doing things they never did. Two neural networks compete: a generator creates synthetic video frames while a discriminator tries to distinguish real from synthetic. The process involves facial mapping, emotion synthesis, and speech generation. Deepfakes are novel in their technical sophistication and their accessibility: they can be created with limited expertise using open-source software and freely available training data (e.g., celebrity photos from the internet).

Benefits of Deepfake Technology¶

Despite their risks, deepfakes have positive applications:

Entertainment and media: Film production, actor voice dubbing for accessibility, realistic special effects, creating content in multiple languages without dubbed audio
Healthcare: Therapeutic applications (helping people with Alzheimer's interact with digital recreations of loved ones)
Education: Multilingual content delivery, avatar creation for distance learning
Accessibility: Digital doubles enabling broader audience engagement
Commerce: Virtual fashion try-ons, product visualization

Who Produces Deepfakes?¶

Four major actor categories:

Hobbyist communities: Amateur users experimenting with deepfake tools; often treat them as entertainment or joke content; hosted on platforms like Reddit before moderation tightened
Political actors: Foreign governments and activists using deepfakes in disinformation campaigns to manipulate elections and undermine institutional trust
Malevolent actors: Fraudsters and criminals using deepfakes for revenge porn, blackmail, impersonation for financial fraud, or identity theft
Legitimate actors: Entertainment companies, filmmakers, broadcasters, and researchers developing beneficial uses

Threats of Deepfakes¶

Deepfakes pose multiple threats:

Misinformation: Journalists struggle to filter real from fake; false videos spread rapidly on social media and can influence public opinion within hours
Election integrity: Deepfakes of political leaders could spread during campaigns, creating uncertainty about authentic statements and influencing voter behavior
Trust erosion: People increasingly doubt anything they watch online, including authentic videos (the "liar's dividend")
National security: Deepfakes could be weaponized for propaganda, election interference, or coordination of hostile actions between nation-states
Cybersecurity: Deepfakes enable social engineering (impersonating executives for wire transfer fraud) and can be used to bypass biometric authentication systems
Non-consensual intimate imagery: Synthetic sexual videos used to harass, abuse, and blackmail (predominantly targeting women)

Current Examples¶

Political: Obama/Peele deepfake (2018), Tom Cruise deepfakes on TikTok, deepfakes of politicians in India/Pakistan conflicts
Celebrity/entertainment: Deepfakes of celebrities in movies they never starred in, synthetic performances using deceased actors' likenesses
Financial fraud: Deepfake audio of executives authorizing wire transfers; criminals using voice synthesis to impersonate leadership
Revenge and harassment: Non-consensual intimate deepfakes, primarily targeting women; used for harassment and blackmail

Methods to Combat Deepfakes¶

Westerlund identifies four complementary strategies:

Legislation and Regulation¶

Some jurisdictions have begun criminalizing deepfakes: - U.S. state laws targeting revenge porn and identity fraud (e.g., Virginia's law on false-created media) - Proposed national regulatory frameworks addressing privacy, copyright, and ownership of facial imagery - Challenges: balancing free speech with prohibition; determining liability between creators and platforms

Corporate Policy and Voluntary Action¶

Platform content moderation policies (YouTube, Facebook, TikTok, Reddit now flag or remove deepfake content)
Proactive enforcement: Facebook and other platforms collaborate with fact-checkers to identify and remove deepfakes before viral spread
Transparency initiatives: Instagram, TikTok, and Twitter label AI-generated content

Education and Media Literacy¶

Training journalists to identify technical artifacts of deepfakes
Teaching critical thinking skills and digital literacy to the general public
Raising awareness among older populations and less tech-savvy groups about deepfake risks
Educational interventions showing that labeled deepfakes are less likely to be believed

Anti-Deepfake Technology¶

Detection methods fall into several categories:

Frequency-domain analysis: Deepfakes exhibit characteristic artifacts in Fourier space due to GAN compression
Behavioral inconsistencies: Unnatural eye blinks, facial expression sequences, or head pose trajectories
Audio-visual synchronization: Lip-sync mismatches or timing inconsistencies between mouth and speech
Forensic analysis: Detection of camera noise patterns, sensor artifacts, lighting inconsistencies
Biometric markers: Photo Response Non-Uniformity (PRNU) patterns in camera sensors; blood-flow changes indicative of naturally breathing faces

Challenges: detection is an arms race—as detection methods improve, generative techniques advance to evade them.

Connections¶

Deepfakes topic page — comprehensive overview of synthetic video technology, technical mechanisms, and detection methods
Vaccari & Chadwick (2020) — empirical evidence that deepfakes increase uncertainty and erode trust in news media
Mirsky (2020) — comprehensive survey of deepfake creation and detection methodologies
Rana et al. (2022) — systematic literature review of deepfake detection approaches
Synthetic media — broader category encompassing deepfakes plus AI-generated images, text, and audio
Misinformation spread — how deepfakes propagate through social networks
Content moderation — platform policies for detecting and removing deepfakes

Notes¶

This review's strength is its accessibility and breadth: it synthesizes information from news media (84 articles) rather than just academic research, capturing how deepfakes are discussed in public discourse. The four-part framework (benefits, threats, examples, defenses) is pedagogically clear and useful for policymakers and technologists.

Limitations: The empirical data is drawn entirely from news articles, which may overrepresent sensational cases (celebrity deepfakes, political threats) relative to the actual distribution of deepfake creation. The paper was written in 2019, before the maturation of diffusion-model-based synthesis and before widespread deployment of detection systems at scale. Some technical details on detection (especially audio-visual synchronization analysis) have been superseded by more recent work.

The finding that "no single approach is sufficient" has proven prescient: five years later, deepfake detection remains unsolved at platform scale, legislation has progressed slowly, and public literacy remains low relative to technical sophistication.