Cyberpediaby Brain

Spotting a deepfake by sight and sound

A synthetic face or voice can imitate an executive or a colleague. What is established, and why to verify through another channel.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4Inside the attacker's head
  5. 5What you see on screen
  6. 6Under the hood
  7. 7Case files
  8. 8Sources
  9. 9Cite this article

In 30 seconds

Where it happens
A phone call or other voice communication[1], or a message sent through a messaging service[5], with an identity borrowed from someone trusted.
What the attacker wants
To gain access to systems by phone[1].
The reflex
Be wary of callers and verify their identity[1]. For this entry, that means confirming the request through a different channel from the one it came by, without trusting the face or the voice.
If it has already happened
Tell the company's security team without delay, with the time of the call and the channel used.
Article map

What it is

A deepfake is an image, a video or a voice fabricated by an artificial intelligence model to resemble a real person. For an employee, the danger is that a familiar face or voice is used to push a request through: an "executive" demanding an urgent transfer, a "colleague" asking for access.

This entry describes what the victim perceives, how such content is made, and why the eye and the ear are not enough to decide.

Straight answers

Can a deepfake be recognized with the naked eye?

Under certain conditions, generation techniques produce "highly realistic" facial images and videos[2]. NIST, for its part, devotes work to evaluating analytic systems against AI-generated image deepfakes[3].

Why do visible flaws disappear?

Generative models keep improving. A survey notes that generative adversarial networks (GANs) produced impressive results, and that diffusion models then revived research[2]. The same survey states that detection techniques are also evolving continuously[2].

What do the authorities say about the threat?

On September 12, 2023, the NSA, the FBI and CISA published an information sheet, "Contextualizing Deepfake Threats to Organizations". It presents the threats, techniques and trends of synthetic media, including deepfakes, and recommends that organizations look in it for measures to prepare, identify, defend and respond[4]. The agencies write that these threats have "exponentially increased"[4].

What if the request comes from a boss?

Verify through another channel. MITRE recommends training users to be wary of callers and to verify their identity[1]. In practice: hang up, then call the person back on a number already known, or reach them through internal messaging.

Inside the attacker's head

The attacker wants someone to act. MITRE describes "spearphishing voice" as a variant of targeted phishing that pushes a user to give access to systems through a phone call or other voice communication[1]. The method often relies on social engineering: posing as a trusted source and creating a sense of urgency or alarm in the person[1].

A synthetic voice or face adds a layer to this scenario: the other party is no longer only someone claiming to be the executive, but has the executive's appearance or tone of voice.

The attacker can also use a trust-building step. MITRE describes attackers who create fake social media accounts, start conversations with employees, then send links or attachments through a service not controlled by the company, such as a personal email account read on a work computer[5]. These services are more likely to have a less strict security policy than a company's[5].

What you see on screen

The victim sees or hears a person they believe they know. The request shares with other phishing a trusted identity and urgency[1].

Here is an illustration, neutralized, of what such a call can contain. It is invented for the example and reproduces no real case.

Excerpt
Illustration (fictional case, address neutralized)
Call from [voice imitating an executive]:
[request to process an urgent and confidential payment,
 request to tell no one, then to log in at
 hxxps://urgent-access[.]example[.]invalid to confirm]

MITRE also notes that voice phishing can be combined with the generation of multifactor authentication requests, to lead the victim to disclose their MFA credentials or to accept authentication requests[1].

Under the hoodIT teams

Two families of models recur.

Generative adversarial networks, or GANs, were proposed in a paper submitted on June 10, 2014. The principle: two models are trained together, a generator that captures the distribution of the training data and a discriminator that estimates whether a sample comes from that data or from the generator[6]. The generator's training aims to maximize the probability that the discriminator makes a mistake[6]. Diffusion models followed. A paper submitted on June 19, 2020 presents high-quality image synthesis results with denoising diffusion probabilistic models, and reports, on the 256x256 LSUN dataset, a sample quality similar to that of ProgressiveGAN[7].

The survey on deepfakes distinguishes four fields: face swapping, expression reconstruction (face reenactment), talking face generation and facial attribute editing, to which is added forgery detection[2].

On the defense side, for fraudulent calls MITRE proposes monitoring the call logs of work devices for unusual or unauthorized numbers, then correlating them with system events that follow, such as the execution of a remote management tool[1].

Case filesReal cases

The two cases below are calls from people posing as IT support, with no indication of a synthetic voice. They show the scenario that deepfakes can strengthen.

Scattered Spider: fake IT staff on the phone

MITRE reports that, during campaign C0027, the Scattered Spider group posed as legitimate IT personnel on calls, to get victims to download a remote monitoring and management (RMM) tool that would let it take control of their workstation[1].

Storm-1811: "technical support" calls

MITRE states that Storm-1811 made voice calls posing as IT support, to prompt users to download and run scripts and tools serving as initial access[1]. Microsoft Threat Intelligence published an article on May 15, 2024 on the misuse of Quick Assist in social engineering attacks leading to ransomware[1].

Sources

  1. MITRE ATT&CK, “Phishing: Spearphishing Voice, Sub-technique T1566.004 - Enterprise”. https://attack.mitre.org/techniques/T1566/004/ (accessed 10 November 2026)
  2. arXiv, “Deepfake Generation and Detection: A Benchmark and Survey”, 26 March 2024. https://arxiv.org/abs/2403.17881 (accessed 10 November 2026)
  3. NIST, “Guardians of Forensic Evidence: Evaluating Analytic Systems Against AI-Generated Deepfakes”, 27 January 2025. https://www.nist.gov/publications/guardians-forensic-evidence-evaluating-analytic-systems-against-ai-generated-deepfakes (accessed 10 November 2026)
  4. CISA, “NSA, FBI, and CISA Release Cybersecurity Information Sheet on Deepfake Threats”, 12 September 2023. https://www.cisa.gov/news-events/alerts/2023/09/12/nsa-fbi-and-cisa-release-cybersecurity-information-sheet-deepfake-threats (accessed 10 November 2026)
  5. MITRE ATT&CK, “Phishing: Spearphishing via Service, Sub-technique T1566.003 - Enterprise”. https://attack.mitre.org/techniques/T1566/003/ (accessed 10 November 2026)
  6. arXiv, “Generative Adversarial Networks”, 10 June 2014. https://arxiv.org/abs/1406.2661 (accessed 10 November 2026)
  7. arXiv, “Denoising Diffusion Probabilistic Models”, 19 June 2020. https://arxiv.org/abs/2006.11239 (accessed 10 November 2026)